Pith. sign in

REVIEW 4 major objections 6 minor 300 references

This thesis argues that reinforcement learning is objective-driven adaptation: in two-player zero-sum games, deep Q-learning with Nash-equilibrium targets yields policies that resist adversarial exploitation, and the same principle extends

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 17:38 UTC pith:ZK2JORBQ

load-bearing objection A thesis compiling already-published RL work with a genuinely useful benchmark, but the non-exploitability claim is overstated and half the chapters are missing. the 4 major comments →

arxiv 2607.17560 v1 pith:ZK2JORBQ submitted 2026-07-20 cs.AI

Reinforcement Learning: From Algorithms To Foundation Models

Zihan Ding This is my paper
classification cs.AI
keywords reinforcement learningNash equilibriumzero-sum Markov gamesnon-exploitabilityworld modelsdiffusion modelsfoundation modelsmulti-agent RL
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This doctoral thesis argues that reinforcement learning is best understood as objective-driven adaptation, and that the same principle underlies both equilibrium computation in games and planning with generative foundation models. The load-bearing claim of the game-theoretic part is that in two-player zero-sum Markov games, a Nash equilibrium strategy is the natural non-exploitable solution: playing it makes one immune to a best-responding adversary. The thesis proposes Nash-DQN and Nash-DQN-Exploiter, which combine deep Q-learning with a Nash-solving subroutine, and reports that the resulting policies achieve non-positive exploiter rewards in five video-game environments, outperforming self-play and policy-space methods. The second half argues that pretrained generative models and learned world models act as structured priors, giving diffusion-based world models, consistency models as policy classes, and memory-augmented video world models. A sympathetic reader comes away with a unified picture of RL as the study of how incentives, environment models, and expressive policy classes combine in sequential decision making.

Core claim

The thesis's central discovery is that a deep Q-network trained with Nash-equilibrium targets can approximate non-exploitable policies in two-player zero-sum Markov games with high-dimensional observations. Concretely, Nash-DQN replaces the single-agent max in DQN's target with the value of a Nash equilibrium of the Q-matrix at the next state, solving a small normal-form game at each update; Nash-DQN-Exploiter additionally trains an adversary that best-responds to the main policy to drive exploration. In tabular environments the method reduces to Nash value iteration, which provably finds approximate Nash equilibria with polynomial sample complexity, and in video games the learned policies w

What carries the argument

The central mechanism of the game-theoretic part is the Nash Bellman optimality update: at each state, the agent solves a normal-form two-player zero-sum game whose payoff matrix is the Q-function over joint actions, and uses the resulting Nash value as the target for the Q-network. This converts single-agent DQN's max-operator into a minimax operator, which is what carries the non-exploitability guarantee. The tabular version, Nash value iteration, is proven to find approximate Nash equilibria in polynomial steps, and the exploiter variant adds a separate value network that tracks the value of the main policy against its best response, providing targeted exploration. In the foundation-model

Load-bearing premise

The load-bearing premise is that the DQN or PPO exploiter used to measure non-exploitability is strong enough to approximate the true best response; if a stronger adversary exists, a policy that looks non-exploitable could still be beaten.

What would settle it

Train a substantially stronger exploiter—e.g., a PPO agent with many more timesteps, a policy trained via game-tree search, or a human expert—against the Nash-DQN policy in Pong or Boxing and measure its expected reward; a positive reward for the stronger adversary would refute the claim that the learned policy is non-exploitable. The thesis itself partially does this: in FightLadder, a PPO exploiter and human players systematically defeat all trained baselines, showing none achieve Nash-level robustness.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If Nash-DQN's reported non-exploitability holds, equilibrium-based deep RL becomes a viable route to robust policies in zero-sum games, since an equilibrium policy cannot lose to any adversary that merely best-responds.
  • The reduction to Nash value iteration in tabular settings implies the practical algorithms inherit a polynomial convergence guarantee, at least when function approximation is not in the way.
  • Modelling multi-agent load balancing as a Markov potential game with variance-based fairness means independent, decentralized learners can reach Nash equilibria that minimize makespan without centralised coordination.
  • If diffusion world models provide accurate long-horizon predictions, model-based RL with them should outperform one-step dynamics models in sample efficiency and planning horizons, as the thesis's experiments indicate.
  • Consistency models used as policies show that generative-model advances can directly enrich RL action spaces, allowing multi-step generation to be collapsed into single-step control.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The non-exploitability claim is only as strong as the adversary used to measure it; the thesis itself reports a DQN exploiter cannot fully best-respond to an oracle Nash policy (exploitability about -0.03), so a stronger adversary, say one with more compute or a search-based planner, could in principle crack policies that pass the current test.
  • If the Nash-DQN approach scales, one could extend it to partial-information games by coupling the Nash target with a belief state or recurrence, since the current method assumes full state observability.
  • The Markov-potential-game formulation of load balancing suggests a template: any system objective that can be written as a variance-based fairness per agent admits decentralized learning with equilibrium guarantees, a pattern applicable to other resource-allocation domains such as traffic routing or job scheduling.
  • In the foundation-model part, the implied next step is to treat a single world model as both a simulator for policy optimization and a target for fine-tuning, so that an agent can improve its model and its policy in a closed loop, a direction the thesis gestures at but does not fully develop.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The thesis presents a two-part dissertation on reinforcement learning. Part I studies game-theoretic RL: Chapter 3 proposes Nash-DQN and Nash-DQN-Exploiter for two-player zero-sum Markov games and reports tabular experiments with exact exploitability as well as Atari/SlimeVolley experiments in which exploitability is estimated by training a DQN exploiter; Chapter 4 presents the FightLadder fighting-game benchmark with PPO-based baselines, Elo ratings, and exploitability evaluations; Chapter 5 formulates network load balancing as a Markov potential game with variance-based fairness. Part II collects work on foundation-model RL: diffusion world models, consistency models as policies, RL for few-step video generation, a video world model with retrieval augmentation, and a world model with memory. The unifying thesis is that RL can be viewed as objective-driven adaptation across strategic games and generative world models.

Significance. The thesis contains several potentially valuable contributions: the tabular comparison of Nash-VI/Nash-VI-Exploiter against self-play, fictitious play, and double-oracle baselines is a clean proof-of-concept; the FightLadder benchmark, including open-source baselines and Elo/exploitability protocols, is a practical resource for competitive MARL; the Markov-potential-game formulation of load balancing with variance-based fairness is a useful formalization. The generated-model chapters report substantial empirical results on diffusion world models, consistency-model policies, video generation, and memory-augmented world models. Had the central Chapter 3 claim—that Nash-DQN finds non-exploitable policies in zero-sum video games—been fully supported, it would be a notable result. However, as argued below, the deep-network non-exploitability claim is not established by the reported measurements.

major comments (4)
  1. [Sec. 3.5.3, Eq. (3.8), Table 3.1, Table 3.2] The central claim that Nash-DQN finds non-exploitable strategies is not supported by the reported experimental evidence. The quantity reported as 'approximate exploitability' is the reward of a DQN exploiter trained against a frozen policy, not the true best-response value used in Eq. (3.8). Table 3.1 calibrates this same DQN exploiter against the ground-truth Oracle Nash policy and obtains -0.027 (Env I) and -0.082 (Env II) instead of 0. The measured quantity is therefore a lower bound on true exploitability, biased in the direction of non-exploitability. Non-positive entries in Table 3.2 are compatible with strictly positive true exploitability; they only show that this particular DQN exploiter could not find a profitable deviation. The sentence in Sec. 3.5.3 that non-positive rewards 'demonstrate the non-exploitability' should be weakened to 'were not exploited by the DQN exploiter us
  2. [Sec. 3.4.3, Theorem 30, Sec. 3.7.9] The theoretical justification covers the tabular, optimistic versions of Nash-VI and Nash-VI-Exploiter, not the deep epsilon-greedy Nash-DQN used in the video-game experiments. The reduction in Sec. 3.7.9 only converts published episodic tabular results to the infinite-horizon discounted tabular setting by truncation and state augmentation. It does not address function approximation, minibatch SGD, target networks, or epsilon-greedy exploration. Thus Theorem 30 cannot be invoked to support the deep-network non-exploitability claim; that claim rests entirely on the biased measurement described in the previous comment. The paper should explicitly separate the tabular proof-of-concept from the heuristic deep-learning transfer.
  3. [Sec. 3.5.3, Table 3.2] For each method and environment, Table 3.2 reports the best-performing model over five seeds rather than the average or all seed-wise results. Best-seed selection biases the comparison in favor of the proposed method and makes the numeric values not representative of expected performance. The exploitation curves in Fig. 3.3 are informative, but the table should report mean and standard deviation (or per-seed values) for the central experimental claim.
  4. [Sec. 4.7.3, Table 4.3] In the FightLadder benchmark results, a PPO exploiter eventually beats policies from all baselines with win rate greater than 0.5, and human players using simple strategies defeat the trained policies. This is consistent with the first major comment: 'exploitability' as measured by a fixed RL exploiter is only a lower bound relative to that exploiter. The benchmark should define the metric as 'win rate against the given exploiter' rather than calling it 'exploitability' in the game-theoretic sense, unless a calibrated or exact best-response oracle is available.
minor comments (6)
  1. [Contents/Ch. 2] Typo: 'Chaper 10' should be 'Chapter 10'.
  2. [Sec. 3.5.1, Table 3.1] The acronym 'NSFP' appears in Sec. 3.5.1 and Table 3.1; it should be 'NFSP' (Neural Fictitious Self-Play) for consistency with the rest of the text.
  3. [Sec. 1.2.2] In the paragraph after Eq. (1.4), the phrase 'in this infinite-horizon MDP' is accidentally duplicated.
  4. [Sec. 3.7.9] The section is titled 'Proof of Theorem 30' but it is a reduction to results in [Liu et al., 2021; Jin et al., 2021b]. This is fine for a thesis, but the title should say 'Proof sketch' or 'Reduction to known tabular results' to avoid implying a fully self-contained proof.
  5. [Sec. 4.4.5] The sentence 'While it is less efficient than FightLadder is the PettingZoo Atari' is grammatically incomplete; it should read 'While it is less efficient than PettingZoo Atari' or similar.
  6. [General] Several cross-references are left as 'Sec. Sec.??' (e.g., Sec. 3.5.3, Sec. 5.2). These placeholders should be resolved in the final version.

Circularity Check

0 steps flagged

No significant circularity: the thesis's central derivations are reductions to prior published theory, not definitions in terms of the target claims.

full rationale

The thesis does not exhibit a circular derivation chain. In Ch. 3, Nash-DQN is explicitly presented as a deep approximation of the tabular Nash-VI algorithm, and the convergence guarantee (Theorem 30) is imported from prior published results [Liu et al., 2021; Jin et al., 2021b]; Section 3.7.9 only supplies the episodic-to-discounted truncation on top of those theorems. This is a standard reliance on external theory rather than a self-definitional reduction. The measured 'exploitability' is an empirical quantity obtained by training a DQN/PPO exploiter; Table 3.1 and Ch. 4.5 explicitly acknowledge that this is a lower bound ('limited to the capability of deep RL algorithms in usage'), which is a validity caveat, not a circularity. The reducing arguments in Sec. 3.7.8 (specializing network to a table, full batch, many GD steps, immediate target update) are concrete and checkable. No equation is defined in terms of the quantity it is claimed to predict, no fitted parameter is renamed as a prediction, and the self-citations (to the author's and collaborators' prior work) are load-bearing only in the way any citation to published theorems is load-bearing; they are not used to forbid alternatives or to smuggle in an ansatz. Therefore no circular step meets the required evidentiary threshold.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The thesis's theoretical core leans on same-group cited results (Liu et al., Jin et al.) and on hand-designed rewards (VBF; FightLadder dense rewards) rather than on new ontological entities. No new particles, forces, dimensions, or ledger entries are introduced. The 'exploiter' networks, the 'Nash subroutine', and the MPG potential function are algorithmic constructs, not entities claimed to exist independently, so they are not catalogued as invented entities. The main non-catalogued construct is the operationally-defined 'exploitability measured by a DQN/PPO exploiter', which is a measurement surrogate whose weakness is documented by the negative Oracle Nash exploitability.

free parameters (4)
  • Dense reward weights lambda per FightLadder game = lambda=3 (SF2, FF2, MK); lambda=1 (SF3, KOF97)
    Hand-chosen in Ch 4.8.1 to tune aggressiveness; benchmark results depend on this shaping because baseline agents cannot learn from sparse rewards (Ch 4.4.4).
  • Frames-per-step and stacked frames (FightLadder) = 8 frames/step for SF2/FF2/MK, 3 for SF3/KOF97; stacks 12/12/12/9/9
    Configuration choices (Table 4.4) defining effective action timing and observation history, not derived from the problem.
  • Exploitability smoothing window and best-seed selection = not stated precisely
    Tables 3.1/3.2 smooth over neighboring episodes and report the best-performing seed per environment (Ch 3.5.2-3.5.3); these measurement choices directly set the reported exploitability values.
  • Nash-DQN exploration rate and target-update intervals = epsilon=0.5 in tabular runs (Sec 3.5.2)
    Standard DQN hyperparameters, with no sensitivity analysis shown; the tabular algorithm used is the epsilon-greedy version, not the optimistic version that Theorem 30 covers.
axioms (5)
  • standard math Minimax theorem, existence of Nash equilibrium in finite zero-sum games, Bellman operator contraction (Banach fixed point)
    Ch 1.2.3 / 1.3.1: the background the zero-sum game chapters build upon.
  • domain assumption Polynomial sample complexity of optimistic Nash-VI and Golf-with-Exploiter in tabular Markov games (Liu et al. 2021; Jin et al. 2021b)
    Theorem 30 is lifted from these cited papers; the same-group provenance is documented in the circularity rationale. The thesis adds only the episodic-to-discounted truncation argument (Sec 3.7.9).
  • domain assumption Maximizing variance-based fairness is sufficient for minimizing makespan (Prop 37; Prop 39 for product-based fairness)
    Central link between the VBF reward and the load-balancing objective; proof deferred to Sec 5.6, whose content is absent from the provided text.
  • domain assumption The synchronous load-balancing model (Alg 15: task arrivals w_i(t), capacity v_j per step, per-LB assignment alpha_ij) captures real data-center load balancers
    Ch 5.1 motivates from real DCs, but all theoretical claims (Theorem 43, Propositions 37/39) are for this idealized model with simultaneous per-step arrivals and fixed capacities.
  • domain assumption Truncating Atari episodes to 300 steps and using RAM observations preserves the zero-sum game structure relevant to non-exploitability
    Ch 3.5.3 experiment setup; truncation changes game dynamics, and the thesis provides no control for this when claiming near-Nash behavior.

pith-pipeline@v1.3.0-alltime-deepseek · 5202 in / 5073 out tokens · 215557 ms · 2026-08-01T17:38:57.161505+00:00 · methodology

0 comments
read the original abstract

Reinforcement learning (RL) provides a framework for sequential decision making under explicit objectives. In its classical form, RL studies how an agent should act to maximise long-term reward in a dynamic environment. In richer settings, the problem extends beyond a single agent and fixed environment: intelligent behavior may require strategic interaction, adaptation to uncertainty, and reasoning over high-dimensional worlds. This thesis studies RL from two perspectives: algorithms in games and RL in the era of foundation models. The first part focuses on multi-agent RL in games. It examines how incentives, policies, and equilibrium concepts interact in competitive and general-sum environments, spanning two-player zero-sum games, large-scale video games, and multi-player settings with general structure. These works investigate learning in multi-agent systems and the behavior of RL methods in interactive environments. The second part studies RL with generative and foundation models, motivated by the idea that prior knowledge can enrich sequential decision making. Pretrained generative models and learned world models serve as representation tools and structured priors for planning, control, and policy optimization. The thesis develops diffusion-based world models, investigates RL for efficient video generation, explores generative models as policy classes, and studies interactive video world models in which actions shape future observations. It also addresses long-horizon modeling through architectures with memory. Together, these contributions present a unified view of RL as objective-driven adaptation in complex sequential domains. From strategic games to generative world models, the thesis highlights how RL connects decision making, environment modeling, and emerging foundation-model capabilities, offering a broader perspective on the principles underlying intelligent behavior.

Figures

Figures reproduced from arXiv: 2607.17560 by Zihan Ding.

Figure 1.1
Figure 1.1. Figure 1.1: A general reinforcement learning system: the agent interacts with the envi [PITH_FULL_IMAGE:figures/full_fig_p018_1_1.png] view at source ↗
Figure 3.1
Figure 3.1. Figure 3.1: Screen shots of the six two-player video games. [PITH_FULL_IMAGE:figures/full_fig_p046_3_1.png] view at source ↗
Figure 3.2
Figure 3.2. Figure 3.2: Tabular case experiments on two randomly generated Markov games. [PITH_FULL_IMAGE:figures/full_fig_p059_3_2.png] view at source ↗
Figure 3.3
Figure 3.3. Figure 3.3: The exploiter learning curves for exploitation tests on six two-player zero-sum [PITH_FULL_IMAGE:figures/full_fig_p061_3_3.png] view at source ↗
Figure 3
Figure 3. Figure 3: and Table 3.2 show the exploitation results of all algorithms and baselines. [PITH_FULL_IMAGE:figures/full_fig_p061_3.png] view at source ↗
Figure 3.4
Figure 3.4. Figure 3.4: The key frames in Boxing exploitation test: (a-c) shows a sub-optimal model exploited by the exploiter. (d) shows our proposed algorithms learn hard-to-exploit policy robust against the turning-around strategy of the exploiter. This shows the challenge for finding approximate Nash equilibrium strategies for these games. Nash-DQN demonstrates significant advantages over other methods across all six games.… view at source ↗
Figure 3.5
Figure 3.5. Figure 3.5: The exploiter learning curves for longer exploitation tests on five video games. [PITH_FULL_IMAGE:figures/full_fig_p063_3_5.png] view at source ↗
Figure 4.1
Figure 4.1. Figure 4.1: FightLadder currently supports various cross-platform video fighting games: [PITH_FULL_IMAGE:figures/full_fig_p077_4_1.png] view at source ↗
Figure 4.2
Figure 4.2. Figure 4.2: Motion and attack action spaces of fighting games. Images are adapted from [PITH_FULL_IMAGE:figures/full_fig_p085_4_2.png] view at source ↗
Figure 4.3
Figure 4.3. Figure 4.3: Example of special moves for character Ryu in StreetFighter II (left to right): [PITH_FULL_IMAGE:figures/full_fig_p085_4_3.png] view at source ↗
Figure 4.4
Figure 4.4. Figure 4.4: The win rate curves and the scheduling distribution bar plot in [PITH_FULL_IMAGE:figures/full_fig_p091_4_4.png] view at source ↗
Figure 4.5
Figure 4.5. Figure 4.5: The payoff matrix for each pair of agents at a certain stage of League training. [PITH_FULL_IMAGE:figures/full_fig_p093_4_5.png] view at source ↗
Figure 4.6
Figure 4.6. Figure 4.6: The distribution of Elo ratings for top ten agents from each baseline. [PITH_FULL_IMAGE:figures/full_fig_p094_4_6.png] view at source ↗
Figure 4
Figure 4. Figure 4: , 4.8, and 4.9 report the payoff matrix of policies within the population for [PITH_FULL_IMAGE:figures/full_fig_p100_4.png] view at source ↗
Figure 4.7
Figure 4.7. Figure 4.7: FSP details (training order from top left to bottom right): For FSP, there [PITH_FULL_IMAGE:figures/full_fig_p101_4_7.png] view at source ↗
Figure 4.8
Figure 4.8. Figure 4.8: PSRO details (training order from top left to bottom right): For PSRO, [PITH_FULL_IMAGE:figures/full_fig_p102_4_8.png] view at source ↗
Figure 4.9
Figure 4.9. Figure 4.9: League training details (training order from top left to bottom right): For [PITH_FULL_IMAGE:figures/full_fig_p103_4_9.png] view at source ↗
Figure 4.10
Figure 4.10. Figure 4.10: The Elo rating for the population of agents trained with IPPO algorithm. [PITH_FULL_IMAGE:figures/full_fig_p104_4_10.png] view at source ↗
Figure 4.11
Figure 4.11. Figure 4.11: The Elo rating for the population of agents trained with 2Timescale algo [PITH_FULL_IMAGE:figures/full_fig_p104_4_11.png] view at source ↗
Figure 4.12
Figure 4.12. Figure 4.12: The Elo rating for the population of agents trained with FSP algorithm. [PITH_FULL_IMAGE:figures/full_fig_p104_4_12.png] view at source ↗
Figure 4.13
Figure 4.13. Figure 4.13: The Elo rating for the population of agents trained with PSRO algorithm. [PITH_FULL_IMAGE:figures/full_fig_p105_4_13.png] view at source ↗
Figure 4.14
Figure 4.14. Figure 4.14: The Elo rating for the population of agents trained with League training. [PITH_FULL_IMAGE:figures/full_fig_p105_4_14.png] view at source ↗
Figure 5.1
Figure 5.1. Figure 5.1: Network load balancing. Last but not least, given the low-latency and high-throughput constraints in the dis￾tributed networking setup, the interactive training procedure of RL models and the centralised-training-decentralised-execution (CTDE) scheme [Foerster et al., 2018] can incur additional communication and management overhead. In this paper, we study the network load balancing problem in multi-agen… view at source ↗
Figure 5.2
Figure 5.2. Figure 5.2: Existing network load balancing algorithms are sub-optimal under real-world [PITH_FULL_IMAGE:figures/full_fig_p109_5_2.png] view at source ↗
Figure 5.3
Figure 5.3. Figure 5.3: Overview of the proposed distributed MARL framework for network LB. [PITH_FULL_IMAGE:figures/full_fig_p116_5_3.png] view at source ↗
Figure 5.4
Figure 5.4. Figure 5.4: Experimental results show that the proposed distributed RL framework us [PITH_FULL_IMAGE:figures/full_fig_p119_5_4.png] view at source ↗
Figure 6.1
Figure 6.1. Figure 6.1: The return of TD3+BC trained us￾ing diffusion world model and one-step dynamics model. World models are foundational blocks of AI systems to perform planning and reasoning [Ha and Schmidhuber, 2018b]. They serve as simulators of real environ￾ments that predict the future outcome of certain actions will produce, and policies can be derived from them. Representative example usages of them in model-based re… view at source ↗
Figure 6.2
Figure 6.2. Figure 6.2: Performances of Algorithm 18 with DWM and one-step models, using different [PITH_FULL_IMAGE:figures/full_fig_p144_6_2.png] view at source ↗
Figure 6.3
Figure 6.3. Figure 6.3: Average observation and reward prediction errors (across 9 tasks and simu [PITH_FULL_IMAGE:figures/full_fig_p167_6_3.png] view at source ↗
Figure 6.4
Figure 6.4. Figure 6.4: Average observation and reward prediction errors (across 9 tasks and simula [PITH_FULL_IMAGE:figures/full_fig_p167_6_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: reports the return of DWM-IQL and DWM-TD3BC across 3 tasks, with dif [PITH_FULL_IMAGE:figures/full_fig_p171_6.png] view at source ↗
Figure 6.5
Figure 6.5. Figure 6.5: Comparison of DWM methods using different evaluation RTG values (dis [PITH_FULL_IMAGE:figures/full_fig_p171_6_5.png] view at source ↗
Figure 6.6
Figure 6.6. Figure 6.6: Normalized discounted returns for each environment. [PITH_FULL_IMAGE:figures/full_fig_p172_6_6.png] view at source ↗
Figure 6.7
Figure 6.7. Figure 6.7: The breakdown prediction errors of DWM at each prediction timestep with [PITH_FULL_IMAGE:figures/full_fig_p173_6_7.png] view at source ↗
Figure 6.8
Figure 6.8. Figure 6.8: The breakdown prediction errors of DWM at each prediction timestep with [PITH_FULL_IMAGE:figures/full_fig_p174_6_8.png] view at source ↗
Figure 7.1
Figure 7.1. Figure 7.1: Average training time (seconds per epoch) for Consistency-BC and Diffusion-BC across tasks. Results from Tab. 7.1 show the ad￾vantage of using multi-modal policy rep￾resentation for offline RL even only with the BC method. For reference purpose, the values in the brackets allow for on￾line evaluation to achieve the best model selection from the set of trained models, which serve as the maximal possible v… view at source ↗
Figure 7.2
Figure 7.2. Figure 7.2: The average normalized scores and training time versus [PITH_FULL_IMAGE:figures/full_fig_p191_7_2.png] view at source ↗
Figure 7.3
Figure 7.3. Figure 7.3: Comparison of variants of Consistency-AC across tasks in offline RL setting. [PITH_FULL_IMAGE:figures/full_fig_p192_7_3.png] view at source ↗
Figure 7.4
Figure 7.4. Figure 7.4: Learning curves of Diffusion-QL and Consistency-AC for online RL and offline [PITH_FULL_IMAGE:figures/full_fig_p194_7_4.png] view at source ↗
Figure 7.5
Figure 7.5. Figure 7.5: The training time (left) and inference time (right) versus denoising steps [PITH_FULL_IMAGE:figures/full_fig_p198_7_5.png] view at source ↗
Figure 7.6
Figure 7.6. Figure 7.6: The average training time (hours) for offline-to-online and online training [PITH_FULL_IMAGE:figures/full_fig_p202_7_6.png] view at source ↗
Figure 8.1
Figure 8.1. Figure 8.1: By incorporating variational score distillation, consistency distillation and [PITH_FULL_IMAGE:figures/full_fig_p204_8_1.png] view at source ↗
Figure 8.2
Figure 8.2. Figure 8.2: Method Overview: The few-step generator Gθ is trained to generate high￾quality samples from random noise in latent space, guided by a combination of variational score distillation (VSD), consistency distillation (CD), and latent reward model (LRM) fine-tuning objectives. VSD loss enhances sample quality, albeit with a risk of mode col￾lapse, while CD loss increases sample diversity without compromising g… view at source ↗
Figure 8.3
Figure 8.3. Figure 8.3: Comparison of different reward fine-tuning methods: (1) Direct reward gra [PITH_FULL_IMAGE:figures/full_fig_p212_8_3.png] view at source ↗
Figure 8.4
Figure 8.4. Figure 8.4: Video samples synthesized with our method and SOTA distillation methods [PITH_FULL_IMAGE:figures/full_fig_p218_8_4.png] view at source ↗
Figure 8.5
Figure 8.5. Figure 8.5: Human preference results for our DOLLAR method (with HPSv2) against [PITH_FULL_IMAGE:figures/full_fig_p218_8_5.png] view at source ↗
Figure 8.6
Figure 8.6. Figure 8.6: Compare the generated samples with (first line) and without (second line) [PITH_FULL_IMAGE:figures/full_fig_p220_8_6.png] view at source ↗
Figure 8.7
Figure 8.7. Figure 8.7: Demonstration of the conjugate velocity prediction: relationship of [PITH_FULL_IMAGE:figures/full_fig_p225_8_7.png] view at source ↗
Figure 8.8
Figure 8.8. Figure 8.8: Visualization of samples in training dataset (left) and samples generated with [PITH_FULL_IMAGE:figures/full_fig_p230_8_8.png] view at source ↗
Figure 8
Figure 8. Figure 8: and Fig. 8.10 show the learning curves of latent reward models (LRMs) with two [PITH_FULL_IMAGE:figures/full_fig_p234_8.png] view at source ↗
Figure 8.9
Figure 8.9. Figure 8.9: The learning process of LRM with HPSv2 reward. [PITH_FULL_IMAGE:figures/full_fig_p234_8_9.png] view at source ↗
Figure 8.10
Figure 8.10. Figure 8.10: The learning process of LRM with PickScore reward. [PITH_FULL_IMAGE:figures/full_fig_p235_8_10.png] view at source ↗
Figure 8.11
Figure 8.11. Figure 8.11: Latent reward model fine-tuning process under reward metrics HPSv2 and [PITH_FULL_IMAGE:figures/full_fig_p235_8_11.png] view at source ↗
Figure 8.12
Figure 8.12. Figure 8.12: Reward model fine-tuning process with VSD+DDPO under reward models [PITH_FULL_IMAGE:figures/full_fig_p238_8_12.png] view at source ↗
Figure 9.1
Figure 9.1. Figure 9.1: A world model possesses mem￾ory capabilities and enables faithful long￾term future prediction by maintaining aware￾ness of its environment and generating pre￾dictions based on the current state and ac￾tions. Example is in Minecraft game. Foundational world models capable of sim￾ulating future outcomes based on different actions are crucial for effective planning and decision-making [Watter et al., 2015; … view at source ↗
Figure 9.2
Figure 9.2. Figure 9.2: Overview of our VRAG framework for interactive video generation. The [PITH_FULL_IMAGE:figures/full_fig_p244_9_2.png] view at source ↗
Figure 9.3
Figure 9.3. Figure 9.3: Visual comparison of VRAG with ground truth videos on world coherence [PITH_FULL_IMAGE:figures/full_fig_p250_9_3.png] view at source ↗
Figure 9.4
Figure 9.4. Figure 9.4: Visual comparison of different methods, evaluated [PITH_FULL_IMAGE:figures/full_fig_p251_9_4.png] view at source ↗
Figure 9.5
Figure 9.5. Figure 9.5: SSIM scores over time for differ￾ent methods on world coherence evaluation. Method SSIM Ò PSNR Ò LPIPS Ó DF (window 10) 0.455 16.161 0.509 DF (window 20) 0.466 16.643 0.538 YaRN 0.462 16.567 0.532 History Buffer 0.459 16.922 0.543 Frame Pack 0.421 16.372 0.574 VRAG 0.506 17.097 0.506 [PITH_FULL_IMAGE:figures/full_fig_p253_9_5.png] view at source ↗
Figure 9.6
Figure 9.6. Figure 9.6: Visual comparison of long-term video prediction (1200 frames) across different [PITH_FULL_IMAGE:figures/full_fig_p253_9_6.png] view at source ↗
Figure 9.7
Figure 9.7. Figure 9.7: SSIM scores over time for com￾pounding error evaluation Method SSIM Ò DF (window 10) 0.297 DF (window 20) 0.321 YaRN 0.316 History Buffer 0.188 Neural Memory 0.283 VRAG 0.349 [PITH_FULL_IMAGE:figures/full_fig_p254_9_7.png] view at source ↗
Figure 9.8
Figure 9.8. Figure 9.8: Visualized video frames on RealEstate10K [PITH_FULL_IMAGE:figures/full_fig_p255_9_8.png] view at source ↗
Figure 9.9
Figure 9.9. Figure 9.9: Comparison of SSIM scores over time for VRAG variants. Method SSIM Ò PSNR Ò LPIPS Ó VRAG 0.506 17.097 0.506 VRAG (no training) 0.455 16.670 0.528 VRAG (no memory) 0.436 16.372 0.547 [PITH_FULL_IMAGE:figures/full_fig_p256_9_9.png] view at source ↗
Figure 9.10
Figure 9.10. Figure 9.10: Comparison of SSIM, PSNR, LPIPS, and discriminator metrics. All metrics [PITH_FULL_IMAGE:figures/full_fig_p260_9_10.png] view at source ↗
Figure 9.11
Figure 9.11. Figure 9.11: Comparison of vanilla long-context extension for DF model and YaRN with [PITH_FULL_IMAGE:figures/full_fig_p262_9_11.png] view at source ↗
Figure 9.12
Figure 9.12. Figure 9.12: Comparison of vanilla long-context extension for DF model and YaRN with [PITH_FULL_IMAGE:figures/full_fig_p263_9_12.png] view at source ↗
Figure 9.13
Figure 9.13. Figure 9.13: Comparison of vanilla long-context extension for DF model and YaRN with [PITH_FULL_IMAGE:figures/full_fig_p263_9_13.png] view at source ↗
Figure 9.14
Figure 9.14. Figure 9.14: Visual comparison of vanilla long-context extension for DF model and YaRN. [PITH_FULL_IMAGE:figures/full_fig_p263_9_14.png] view at source ↗
Figure 9.15
Figure 9.15. Figure 9.15: Training Loss Curves 9.6.4 Predicted Global State In the paper, our main experiments are conducted with the access to the ground-truth global state as conditions during training and inference. However, the practical usage may require the global state to be also predicted based on historical states and actions. To ablate this effect, we trained a pose (global state) prediction model that takes the curren… view at source ↗
Figure 10.1
Figure 10.1. Figure 10.1: Training paradigm for Recurrent Autoregressive Diffusion with global mem [PITH_FULL_IMAGE:figures/full_fig_p274_10_1.png] view at source ↗
Figure 10.2
Figure 10.2. Figure 10.2: Recurrent Autoregressive Diffusion model architecture: The RAD model [PITH_FULL_IMAGE:figures/full_fig_p275_10_2.png] view at source ↗
Figure 10.3
Figure 10.3. Figure 10.3: Comparison of effective temporal attention maps for chunk-wise and frame [PITH_FULL_IMAGE:figures/full_fig_p276_10_3.png] view at source ↗
Figure 10.4
Figure 10.4. Figure 10.4: Comparison of standard forward (left) and hidden states prefetching process [PITH_FULL_IMAGE:figures/full_fig_p277_10_4.png] view at source ↗
Figure 10.5
Figure 10.5. Figure 10.5: Visualization results on Minecraft dataset. Frame-wise RNN can effectively [PITH_FULL_IMAGE:figures/full_fig_p282_10_5.png] view at source ↗
Figure 10.6
Figure 10.6. Figure 10.6: Comparison of chunk-wise TTT and frame-wise TTT [PITH_FULL_IMAGE:figures/full_fig_p284_10_6.png] view at source ↗
Figure 10.7
Figure 10.7. Figure 10.7: Training loss curves for frame-wise LSTM with clean frame memory and noised frame memory. Metrics noised clean PSNR Ò 14.70 15.30 SSIM Ò 0.38 0.41 LPIPS Ó 0.52 0.50 [PITH_FULL_IMAGE:figures/full_fig_p285_10_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

300 extracted references · 157 linked inside Pith

  1. [1]

    Neurocomputing , volume=

    Natural actor-critic , author=. Neurocomputing , volume=. 2008 , publisher=

    2008

  2. [2]

    arXiv preprint arXiv:2511.12940 , year=

    Recurrent autoregressive diffusion: Global memory meets local attention , author=. arXiv preprint arXiv:2511.12940 , year=

    Pith/arXiv arXiv

  3. [3]

    International Conference on Machine Learning , pages=

    Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor , author=. International Conference on Machine Learning , pages=

  4. [4]

    Advances in neural information processing systems , pages=

    Actor-critic algorithms , author=. Advances in neural information processing systems , pages=

  5. [5]

    Advances in neural information processing systems , volume=

    Multi-agent actor-critic for mixed cooperative-competitive environments , author=. Advances in neural information processing systems , volume=

  6. [6]

    Mastering the game of Go without human knowledge , volume =

    Silver, David and Schrittwieser, Julian and Antonoglou, Ioannis and Huang, Aja and Guez, Arthur and Hubert, Thomas and Baker, Lucas and Lai, Matthew and Bolton, Adrian and Chen, Yutian and Lillicrap, Timothy and Hui, Fan and Sifre, Laurent and Driessche, George and Graepel, Thore and Hassabis, Demis , year =. Mastering the game of Go without human knowled...

  7. [7]

    International Conference on Learning Representations , year=

    Emergent Tool Use From Multi-Agent Autocurricula , author=. International Conference on Learning Representations , year=

  8. [8]

    , title =

    PKU-Yuan Lab and Tuzhan AI etc. , title =. doi:10.5281/zenodo.10948109 , url =

    doi:10.5281/zenodo.10948109

  9. [9]

    2025 , url=

    MAGI-1: Autoregressive Video Generation at Scale , author=. 2025 , url=

    2025

  10. [10]

    arXiv preprint arXiv:2408.16767 , year=

    Reconx: Reconstruct any scene from sparse views with video diffusion model , author=. arXiv preprint arXiv:2408.16767 , year=

    Pith/arXiv arXiv

  11. [11]

    2025 , eprint=

    WORLDMEM: Long-term Consistent World Simulation with Memory , author=. 2025 , eprint=

    2025

  12. [12]

    arXiv preprint arXiv:2411.02385 , year=

    How far is video generation from world model: A physical law perspective , author=. arXiv preprint arXiv:2411.02385 , year=

    Pith/arXiv arXiv

  13. [13]

    arXiv preprint arXiv:2405.10314 , year=

    Cat3d: Create anything in 3d with multi-view diffusion models , author=. arXiv preprint arXiv:2405.10314 , year=

    Pith/arXiv arXiv

  14. [14]

    2025 , eprint=

    Cosmos-Drive-Dreams: Scalable Synthetic Driving Data Generation with World Foundation Models , author=. 2025 , eprint=

    2025

  15. [15]

    arXiv preprint arXiv:2504.20995 , year=

    TesserAct: Learning 4D Embodied World Models , author=. arXiv preprint arXiv:2504.20995 , year=

    Pith/arXiv arXiv

  16. [16]

    arXiv preprint arXiv:2403.14773 , year=

    Streamingt2v: Consistent, dynamic, and extendable long video generation from text , author=. arXiv preprint arXiv:2403.14773 , year=

    Pith/arXiv arXiv

  17. [17]

    arXiv preprint arXiv:2503.03751 , year=

    Gen3c: 3d-informed world-consistent video generation with precise camera control , author=. arXiv preprint arXiv:2503.03751 , year=

    Pith/arXiv arXiv

  18. [18]

    arXiv preprint arXiv:2503.18945 , year=

    Aether: Geometric-aware unified world modeling , author=. arXiv preprint arXiv:2503.18945 , year=

    Pith/arXiv arXiv

  19. [19]

    Advances in neural information processing systems , volume=

    Mcvd-masked conditional video diffusion for prediction, generation, and interpolation , author=. Advances in neural information processing systems , volume=

  20. [20]

    Advances in Neural Information Processing Systems , volume=

    Elucidating the design space of diffusion-based generative models , author=. Advances in Neural Information Processing Systems , volume=

  21. [21]

    arXiv preprint arXiv:2303.08774 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    Pith/arXiv arXiv

  22. [22]

    arXiv preprint arXiv:2403.11027 , year=

    Reward guided latent consistency distillation , author=. arXiv preprint arXiv:2403.11027 , year=

    Pith/arXiv arXiv

  23. [23]

    arXiv preprint arXiv:2210.02410 , year=

    The vendi score: A diversity evaluation metric for machine learning , author=. arXiv preprint arXiv:2210.02410 , year=

    Pith/arXiv arXiv

  24. [24]

    arXiv preprint arXiv:2207.12598 , year=

    Classifier-free diffusion guidance , author=. arXiv preprint arXiv:2207.12598 , year=

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2408.06072 , year=

    Cogvideox: Text-to-video diffusion models with an expert transformer , author=. arXiv preprint arXiv:2408.06072 , year=

    Pith/arXiv arXiv

  26. [26]

    arXiv preprint arXiv:2304.11277 , year=

    Pytorch fsdp: experiences on scaling fully sharded data parallel , author=. arXiv preprint arXiv:2304.11277 , year=

    Pith/arXiv arXiv

  27. [27]

    arXiv preprint arXiv:2410.13720 , year=

    Movie gen: A cast of media foundation models , author=. arXiv preprint arXiv:2410.13720 , year=

    Pith/arXiv arXiv

  28. [28]

    Advances in Neural Information Processing Systems , volume=

    Pick-a-pic: An open dataset of user preferences for text-to-image generation , author=. Advances in Neural Information Processing Systems , volume=

  29. [29]

    arXiv preprint arXiv:2409.00588 , year=

    Diffusion policy policy optimization , author=. arXiv preprint arXiv:2409.00588 , year=

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2310.04378 , year=

    Latent consistency models: Synthesizing high-resolution images with few-step inference , author=. arXiv preprint arXiv:2310.04378 , year=

    Pith/arXiv arXiv

  31. [31]

    arXiv preprint arXiv:2209.14687 , year=

    Diffusion posterior sampling for general noisy inverse problems , author=. arXiv preprint arXiv:2209.14687 , year=

    Pith/arXiv arXiv

  32. [32]

    arXiv preprint arXiv:2410.05954 , year=

    Pyramidal Flow Matching for Efficient Video Generative Modeling , author=. arXiv preprint arXiv:2410.05954 , year=

    arXiv

  33. [33]

    arXiv preprint arXiv:2305.13301 , year=

    Training diffusion models with reinforcement learning , author=. arXiv preprint arXiv:2305.13301 , year=

    Pith/arXiv arXiv

  34. [34]

    2024 , howpublished =

    Kuaishou , title =. 2024 , howpublished =

    2024

  35. [35]

    arXiv preprint arXiv:2209.14988 , year=

    Dreamfusion: Text-to-3d using 2d diffusion , author=. arXiv preprint arXiv:2209.14988 , year=

    Pith/arXiv arXiv

  36. [36]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  37. [37]

    Advances in Neural Information Processing Systems , volume=

    Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation , author=. Advances in Neural Information Processing Systems , volume=

  38. [38]

    arXiv preprint arXiv:2407.08737 , year=

    Video Diffusion Alignment via Reward Gradients , author=. arXiv preprint arXiv:2407.08737 , year=

    Pith/arXiv arXiv

  39. [39]

    The Twelfth International Conference on Learning Representations , year=

    Instaflow: One step is enough for high-quality diffusion-based text-to-image generation , author=. The Twelfth International Conference on Learning Representations , year=

  40. [40]

    arXiv preprint arXiv:2401.12945 , year=

    Lumiere: A space-time diffusion model for video generation , author=. arXiv preprint arXiv:2401.12945 , year=

    Pith/arXiv arXiv

  41. [41]

    arXiv preprint arXiv:2309.15103 , year=

    Lavie: High-quality video generation with cascaded latent diffusion models , author=. arXiv preprint arXiv:2309.15103 , year=

    Pith/arXiv arXiv

  42. [42]

    arXiv preprint arXiv:1802.05668 , year=

    Model compression via distillation and quantization , author=. arXiv preprint arXiv:1802.05668 , year=

    Pith/arXiv arXiv

  43. [43]

    arXiv preprint arXiv:1312.6114 , year=

    Auto-encoding variational bayes , author=. arXiv preprint arXiv:1312.6114 , year=

    Pith/arXiv arXiv

  44. [44]

    Advances in neural information processing systems , volume=

    Generative modeling by estimating gradients of the data distribution , author=. Advances in neural information processing systems , volume=

  45. [45]

    arXiv preprint arXiv:2406.04324 , year=

    SF-V: Single Forward Video Generation Model , author=. arXiv preprint arXiv:2406.04324 , year=

    Pith/arXiv arXiv

  46. [46]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  47. [47]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Align your latents: High-resolution video synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  48. [48]

    2024 , url =

    Zangwei Zheng and Xiangyu Peng and Tianji Yang and Chenhui Shen and Shenggui Li and Hongxin Liu and Yukun Zhou and Tianyi Li and Yang You , title =. 2024 , url =

    2024

  49. [49]

    Advances in Neural Information Processing Systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in Neural Information Processing Systems , volume=

  50. [50]

    arXiv preprint arXiv:2409.08861 , year=

    Adjoint Matching: Fine-tuning Flow and Diffusion Generative Models with Memoryless Stochastic Optimal Control , author=. arXiv preprint arXiv:2409.08861 , year=

    Pith/arXiv arXiv

  51. [51]

    International Conference on Learning Representations , year=

    Score-based generative modeling through stochastic differential equations , author=. International Conference on Learning Representations , year=

  52. [52]

    International Conference on Machine Learning , pages=

    Deep unsupervised learning using nonequilibrium thermodynamics , author=. International Conference on Machine Learning , pages=

  53. [53]

    arXiv preprint arXiv:2204.03458 , year=

    Video Diffusion Models , author=. arXiv preprint arXiv:2204.03458 , year=

    Pith/arXiv arXiv

  54. [54]

    arXiv preprint arXiv:2210.02303 , year=

    Imagen Video: High Definition Video Generation with Diffusion Models , author=. arXiv preprint arXiv:2210.02303 , year=

    Pith/arXiv arXiv

  55. [55]

    arXiv preprint arXiv:2205.11495 , year=

    Flexible Diffusion Modeling of Long Videos , author=. arXiv preprint arXiv:2205.11495 , year=

    Pith/arXiv arXiv

  56. [56]

    arXiv preprint arXiv:2204.03638 , year=

    Long Video Generation with Time-Agnostic VQGAN and Time-Sensitive Transformer , author=. arXiv preprint arXiv:2204.03638 , year=

    Pith/arXiv arXiv

  57. [57]

    arXiv preprint arXiv:2208.04531 , year=

    Unconditional Video Generation with Latent Flow Diffusion Models , author=. arXiv preprint arXiv:2208.04531 , year=

    Pith/arXiv arXiv

  58. [58]

    arXiv preprint arXiv:2209.14792 , year=

    Make-A-Video: Text-to-Video Generation without Text-Video Data , author=. arXiv preprint arXiv:2209.14792 , year=

    Pith/arXiv arXiv

  59. [59]

    arXiv preprint arXiv:2210.02399 , year=

    Phenaki: Variable Length Video Generation from Open Domain Textual Descriptions , author=. arXiv preprint arXiv:2210.02399 , year=

    Pith/arXiv arXiv

  60. [60]

    arXiv preprint arXiv:2104.10157 , year=

    VideoGPT: Video Generation using VQ-VAE and Transformers , author=. arXiv preprint arXiv:2104.10157 , year=

    Pith/arXiv arXiv

  61. [61]

    arXiv preprint arXiv:2301.06513 , year=

    Dual Diffusion Models for High-Fidelity Video Generation , author=. arXiv preprint arXiv:2301.06513 , year=

    Pith/arXiv arXiv

  62. [62]

    arXiv preprint arXiv:2205.15868 , year=

    CogVideo: Large-scale Pretraining for Text-to-Video Generation with Transformers , author=. arXiv preprint arXiv:2205.15868 , year=

    Pith/arXiv arXiv

  63. [63]

    arXiv preprint arXiv:2302.01327 , year=

    Text2Video-Zero: Zero-Shot Text-to-Video Generation using Pretrained Text-to-Image Diffusion Models , author=. arXiv preprint arXiv:2302.01327 , year=

    Pith/arXiv arXiv

  64. [64]

    arXiv preprint arXiv:2308.06571 , year=

    Modelscope text-to-video technical report , author=. arXiv preprint arXiv:2308.06571 , year=

    Pith/arXiv arXiv

  65. [65]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Videocrafter2: Overcoming data limitations for high-quality video diffusion models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  66. [66]

    arXiv preprint arXiv:2306.09341 , year=

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis , author=. arXiv preprint arXiv:2306.09341 , year=

    Pith/arXiv arXiv

  67. [67]

    arXiv preprint arXiv:2311.15127 , year=

    Stable video diffusion: Scaling latent video diffusion models to large datasets , author=. arXiv preprint arXiv:2311.15127 , year=

    Pith/arXiv arXiv

  68. [68]

    arXiv preprint arXiv:2209.03003 , year=

    Flow straight and fast: Learning to generate and transfer data with rectified flow , author=. arXiv preprint arXiv:2209.03003 , year=

    Pith/arXiv arXiv

  69. [69]

    2024 , archivePrefix=

    HunyuanVideo: A Systematic Framework For Large Video Generative Models , author=. 2024 , archivePrefix=

    2024

  70. [70]

    2025 , eprint=

    Movie Gen: A Cast of Media Foundation Models , author=. 2025 , eprint=

    2025

  71. [71]

    Forty-first international conference on machine learning , year=

    Scaling rectified flow transformers for high-resolution image synthesis , author=. Forty-first international conference on machine learning , year=

  72. [72]

    arXiv preprint arXiv:2210.02747 , year=

    Flow matching for generative modeling , author=. arXiv preprint arXiv:2210.02747 , year=

    Pith/arXiv arXiv

  73. [73]

    arXiv preprint arXiv:2402.00769 , year=

    Animatelcm: Accelerating the animation of personalized diffusion models and adapters with decoupled consistency learning , author=. arXiv preprint arXiv:2402.00769 , year=

    Pith/arXiv arXiv

  74. [74]

    arXiv preprint arXiv:2312.09109 , year=

    Videolcm: Video latent consistency model , author=. arXiv preprint arXiv:2312.09109 , year=

    Pith/arXiv arXiv

  75. [75]

    International Conference on Learning Representations , year=

    Denoising Diffusion Implicit Models , author=. International Conference on Learning Representations , year=

  76. [76]

    arXiv preprint arXiv:2202.00512 , year=

    Progressive Distillation for Fast Sampling of Diffusion Models , author=. arXiv preprint arXiv:2202.00512 , year=

    Pith/arXiv arXiv

  77. [77]

    arXiv preprint arXiv:2106.05237 , year=

    Knowledge Distillation for Generative Models , author=. arXiv preprint arXiv:2106.05237 , year=

    Pith/arXiv arXiv

  78. [78]

    arXiv preprint arXiv:2309.17400 , year=

    Directly fine-tuning diffusion models on differentiable rewards , author=. arXiv preprint arXiv:2309.17400 , year=

    Pith/arXiv arXiv

  79. [79]

    Advances in Neural Information Processing Systems , volume=

    Laion-5b: An open large-scale dataset for training next generation image-text models , author=. Advances in Neural Information Processing Systems , volume=

  80. [80]

    Advances in Neural Information Processing Systems , volume=

    Imagereward: Learning and evaluating human preferences for text-to-image generation , author=. Advances in Neural Information Processing Systems , volume=

Showing first 80 references.