PaperlayerView exact revision on arXiv

Source-linked AI summary

REEL: Learning to Retrieve, Evaluate, and Evolve Visual Memory for Any Long-Horizon Multi-Shot Video Generation

Hang Xu, Guohui Zhang, Xiaoxiao Ma, Hu Yu, Jie Huang, Jiayin Xiao, Lin Song, Haoyang Huang, Nan Duan, Feng Zhao

arXiv:2610.16632v1cs.CVcs.AI

TL;DR

Long-horizon video generation must combine user control with visual consistency across shots, while existing systems often rely on generator-specific interfaces or predefined workflows. REEL learns a visual-memory planner that retrieves and creates evidence, updates memory, and uses adapters to guide different generators; it improves benchmark outcomes on two backends.

  • Problem

    Multi-shot systems often couple memory to generator-specific interfaces or predefined workflows rather than independently training memory policies for cross-shot objectives.

    Source: §1 · p. 2

  • Method

    REEL learns a backend-independent planner to retrieve and inspect references, create missing evidence, review candidates, and update memory, with adapters translating its plans for generators.

    Source: §1 · p. 2

  • Results

    REEL improves ST-Bench cross-shot consistency by 15.1% and 20.7% over No Generated Memory, and REEL-Bench Core-100 by 1.21 and 4.25 points over DirectorAgent on JoyAI-Echo 1.5 and MiniMax-H3, respectively.

    Source: §5 · p. 9

  • Takeaways & Limitations

    The same frozen planner transfers through adapters to four unseen video backends without target-backend fine-tuning.

    Source: §1 · p. 2

    Long-horizon video generation can be misused for misleading content or impersonation; deployment should respect consent and privacy, label synthetic content, and retain generator safety controls.

    Source: ETHICS STATEMENT · p. 10

Abstract

from arXiv · show

Memory-based long-video generation aims to combine user control with long-range visual consistency. We organize this process into three functional components: a planner, an adapter, and a generator. Representative multi-shot systems couple memory to generator-specific interfaces or use predefined agent workflows, rather than independently training a memory policy for cross-shot objectives. We introduce REEL, a general-purpose visual-memory planner. Through a unified tool interface, REEL learns to retrieve references, create missing visual evidence, and update persistent memory. Backend-specific adapters translate its evidence plans into supported conditioning inputs, allowing the same policy to guide different black-box video generators. To train REEL, we construct two high-quality datasets, REEL-SFT-4k and REEL-RL-1k, covering diverse interaction patterns and prompt types. Training proceeds in three stages: supervised fine-tuning, single-shot Group Relative Policy Optimization (GRPO), and multi-shot GRPO. Multi-shot GRPO carries policy-generated memory across consecutive shots, so later outcomes guide earlier reference selection and memory updates. REEL-Bench tests identity preservation and intended state changes under text-only requests and timed reference-image uploads. On JoyAI-Echo~1.5 and MiniMax-H3, REEL improves cross-shot consistency on ST-Bench by 15.1\% and 20.7\%, and Core-100 on our proposed REEL-Bench by 2.41 and 8.32 points, respectively, over the no-memory baseline.

1 INTRODUCTION

Long-horizon multi-shot generation must preserve recurring identities while accommodating requested changes, but existing systems rely on generator-specific interfaces or predefined workflows rather than independently trained cross-shot memory policies. REEL trains a planner to manage visual evidence across shots and evaluates it with new training datasets and a benchmark.

1 INTRODUCTION

  • REEL uses a backend-independent tool interface to retrieve and inspect evidence, create missing references, review candidates, and maintain memory, while adapters map evidence plans to generator inputs.Source: §1 · p. 2
  • REEL-SFT-4k and REEL-RL-1k use VLM-teacher interactions with real generation services; REEL-Bench covers 300 stories and 2,700 shots with text-only requests and timed image uploads.Source: §1 · pp. 2–3
  • Multi-shot GRPO carries policy-produced memory across consecutive requests, letting later outcomes shape earlier reference selection and memory writeback beyond single-shot training.Source: §1 · p. 2
  • REEL leads the compared methods on REEL-Bench Core-100, state compliance, and narrative coherence, and on ST-Bench overall and top-10 cross-shot consistency.Ablations show the largest quality drop when historical retrieval is removed; the frozen planner also transfers to four unseen backends through adapters without target-backend fine-tuning.Source: §1 · p. 2
  • 15.1% and 20.7% higher ST-Bench cross-shot consistency than No Generated Memory, plus 1.21 and 4.25-point REEL-Bench Core-100 gains over DirectorAgent, respectively, on JoyAI-Echo 1.5 and MiniMax-H3.Source: §1 · p. 3

2 RELATED WORK

Prior work addresses long-video continuity, reference-conditioned generation, and visual or persistent memory through temporal methods, generator-internal mechanisms, or task-specific workflows. REEL instead learns visual-memory decisions through a backend-independent planner and uses adapters for generator-specific conditioning.

2 RELATED WORK

  • Long-video methods extend or continue a temporal stream, whereas REEL selects evidence across independently generated, editable shots when entities return beyond recent context.Source: §2 · p. 3
  • Reference-conditioned and story-generation systems preserve subjects across images or clips, while other methods store or route memory inside generators or at shot interfaces.Source: §2 · p. 3
  • Figure 2 depicts shared local resources across shots: solid arrows indicate adjacent-shot context, while dashed arrows indicate writeback and retrieval, including reuse of a Shot 5 reference in Shot 8.Source: Figure 2 · p. 3
  • GroundShot and CoTriSyGen use task-specific orchestration around pretrained generators; REEL learns a visual-memory policy and delegates generator-specific conditioning to adapters.Source: §2 · p. 3
  • Tool-using agents and memory systems learn action patterns, retrieval, storage, or updates, while REEL ties memory decisions to stochastic visual outcomes that can affect later shots.Source: §2 · p. 3
  • Figure 3 traces single-shot inference from requirement analysis and resource queries through evidence creation, backend-specific conditioning, and candidate review.Source: Figure 3 · p. 4

3 METHODOLOGY

REEL separates visual-memory planning from backend adaptation and video synthesis, using persistent resources and tool-mediated evidence plans. The methodology also describes the training data, benchmark, and three-stage planner training scheme.

3.1 PIPELINE

  • The planner submits a backend-agnostic prompt and ranked reference bindings, which an adapter maps to supported inputs for a fixed, black-box video generator.Candidate clips and condition feedback return to the planner for review and possible revision.Source: §3.1 · p. 4
  • Planning uses immutable user assets, a metadata-indexed memory bank, and a subject-linked material registry; selected memories support later reappearances without storing the full video history.The preceding clip’s tail frame is supplied separately for short-range continuity, and visual material must be inspected before use as a condition.Source: §3.1 · p. 4
  • Retrieval, reference creation, candidate review, and memory writeback are conditional tool decisions rather than a fixed execution sequence.The planner can use uploads directly, discard unsuitable references, create missing ones, or finish without writing back.Source: §3.1 · pp. 4–5

3.2 DATA ENGINE

  • REEL-SFT-4k is built from filtered teacher action–observation trajectories, while REEL-RL-1k provides requests and initial resource states for single-shot tasks and contiguous multi-shot episodes.The data engine expands seed stories, uses real image and video generation, and checks tool-call correspondence, legal state transitions, and media provenance.Source: §3.2 · p. 5, Figure 4 · p. 5
  • REEL-Bench contains 300 stories and 2,700 shots, testing identity preservation and evolving instructions under text-only requests and timed reference-image uploads.Methods receive identical requests and scheduled uploads, cannot access future uploads, and maintain memory from their own generated outputs; the training split shares no exact story identifiers with the benchmark.Source: §3.2 · pp. 5–6, Figure 4 · p. 5

3.3 TRAINING SCHEME

  • Training progresses from SFT on demonstrated tool names and arguments to single-shot GRPO, which rewards protocol compliance, evidence-trajectory quality, and generated-video quality.Only planner action tokens receive SFT supervision; the video-quality assessment covers prompt adherence, subject and scene fidelity, motion, and artifacts.Source: §3.3 · p. 6
  • Multi-shot GRPO carries policy-produced memory across contiguous shots, so later identity recovery or obsolete-state propagation affects returns for earlier reference and writeback decisions.Its episode reward combines memory utility for continuity and intended state changes with existing reward components and each shot’s local reward.Source: §3.3 · p. 6

4 EXPERIMENTS

Experiments compare REEL with memory baselines on two benchmarks, ablate training and memory components, and test transfer to unseen backends. REEL improves benchmark results, while ablations and transfer results characterize the contributions and scope of its approach.

4.1 EXPERIMENTAL SETUP

  • Comparisons use JoyAI-Echo 1.5 and MiniMax-H3, matched requests, uploads, settings, and first-attempt seeds, with up to two video candidates and one condition-plan revision per shot.The full jointly trained REEL checkpoint uses Qwen3.5-9B; benchmark metrics are averaged over complete, decodable stories.Source: §4.1 · p. 7

4.2 MAIN RESULTS

  • On Clockwork Skyfall, REEL more closely preserves a requested dodecahedral stone’s geometry and appearance than heuristic policies, using an object reference with preceding-shot context.Figure 5 shows selected final-clip midpoints from eight-shot comparisons under matched requests and upload schedules, with reference-image and video-candidate calls logged.Source: §4.2 · p. 7, Figure 5 · p. 9
  • REEL improves REEL-Bench Core-100 by 1.21 and 4.25 points over DirectorAgent, and ST-Bench overall cross-shot consistency by 15.1% and 20.7% over No Generated Memory.On REEL-Bench it also scores higher in prompt following, state compliance, and narrative coherence on both backends; on ST-Bench it leads in top-10 consistency, while DirectorAgent scores higher in global prompt following.Source: §4.2 · p. 7, Table 1 · p. 8, Table 2 · p. 8

4.3 ABLATION STUDIES

  • Multi-shot GRPO adds 1.2 shots per story and improves Shot and Story by 1.40 and 0.24 points over single-shot GRPO.The base planner averages 0.8 completed shots per story, versus 5.5 after SFT, 7.8 after single-shot GRPO, and 9.0 for REEL.Source: §4.3 · p. 8, Table 3 · p. 9
  • Removing historical retrieval causes the largest ablation drop: 7.59 Shot and 10.36 Story points; disabling creation or review lowers Story by 2.84 or 3.13 points.The creation and review variants retain historical retrieval.Source: §4.3 · p. 8, Table 3 · p. 9
  • The frozen REEL checkpoint transfers through adapters to four unseen backends, with Shot scores from 53.59 to 58.31 and Story scores from 43.03 to 46.29, without target-backend fine-tuning.The target backends contribute no REEL-specific demonstrations or rollouts; adapters map the shared evidence-plan schema to backend-specific conditions.Source: Table 3 · p. 9, §4.3 · p. 9

5 CONCLUSION

REEL is a general-purpose visual-memory planner that reuses one policy across video backends through adapters. It improves benchmark scores over the stated baselines on JoyAI-Echo 1.5 and MiniMax-H3.

5 CONCLUSION

  • REEL learns visual-memory management through a unified tool interface and reuses the same policy across video backends through adapters.The conclusion describes REEL as a general-purpose planner.Source: §5 · p. 9
  • REEL-SFT-4k and REEL-RL-1k support training, while REEL-Bench evaluates identity preservation and state changes under evolving requests and reference uploads.Source: §5 · p. 9
  • REEL improves REEL-Bench Core-100 over DirectorAgent by 1.21 and 4.25 points, and ST-Bench cross-shot consistency over No Generated Memory by 15.1% and 20.7%, respectively.Source: §5 · p. 9

AI USE STATEMENT

The authors report using generative AI across research and manuscript workflows, and accept responsibility for the final work.

AI USE STATEMENT

  • Generative AI tools assisted with experimental design, code implementation, results analysis, manuscript drafting and translation, language editing, and figure preparation.Source: AI USE STATEMENT · p. 10
  • Generative models also supported synthetic data construction, planner-trajectory collection, and VLM-based evaluation.The statement refers to Appendices A and C for these uses.Source: AI USE STATEMENT · p. 10
  • The authors take full responsibility for the final content, including text, claims, code, and reported results.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

The statement identifies third-party artifact restrictions and responsible-use considerations for long-horizon video generation.

ETHICS STATEMENT

  • Use and release of third-party data, models, and other artifacts are subject to applicable licenses, terms of use, and redistribution restrictions.Source: ETHICS STATEMENT · p. 10
  • Long-horizon video generation can be misused to produce misleading content or impersonate individuals.Source: ETHICS STATEMENT · p. 10
  • Deployment should respect consent and privacy when using reference images or voices, identify generated content as synthetic, and retain the underlying generators’ safety controls.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

The reproducibility statement directs readers to descriptions of the method, experiments, data pipeline, and prompts, and describes a planned release subject to licensing and redistribution restrictions.

REPRODUCIBILITY STATEMENT

  • The formulation, planner–adapter–generator interface, and training scheme are described in Section 3.The experimental setup is in Section 4.1; benchmark construction, metrics, and score aggregation are in Appendices C and D.Source: REPRODUCIBILITY STATEMENT · p. 10
  • Appendices A and B document the data pipeline and dataset statistics, while Appendix E covers baselines, generation settings, and training configuration.Appendix G provides the complete system prompts.Source: REPRODUCIBILITY STATEMENT · p. 10
  • The authors plan to release code, planner checkpoints, benchmark inputs, and configurations, subject to applicable licenses and redistribution restrictions.For third-party resources that cannot be redistributed, they plan to provide acquisition instructions.Source: REPRODUCIBILITY STATEMENT · p. 10

A DATA CONSTRUCTION AND EXECUTION PIPELINE

The supplement documents how requests and reference assets are prepared, how planner trajectories are executed and filtered, and how training data is exported while preserving what information is available at each shot.

A.1 STORY INPUTS AND REQUEST DIVERSIFICATION

  • Source stories retain their original shot order; optional diversification varies request presentation and interaction types without changing events or explicit constraints.Eight evidence-input profiles describe what the planner receives, not required tool sequences or quality labels.Source: §A.1 · p. 13

A.2 SHOT-DRIVEN REFERENCE PREPARATION

  • Shot-driven reference preparation plans purpose-specific images, checks generated assets against their intended use, and reveals accepted uploads only at their scheduled shots.Later uploads remain hidden from earlier planning and comparisons.Source: §A.2 · pp. 13–14

A.3 COLLECTING EXECUTABLE PLANNER TRAJECTORIES

  • Each shot begins with a fresh planner conversation, while committed memory, subject records, and accepted prior-shot context persist; visual inspection authorizes retrieved representations.The interface supports retrieval, image creation, plan submission, candidate review, and transactional memory updates; the host and adapter control backend endpoints, seeds, and condition slots.Source: §A.3 · p. 14, Table 4 · p. 14

A.4 FILTERING AND EXPORTING SUPERVISION

  • Review and repair follow runtime checklists and budgets, and durable writeback retains reusable evidence with subject, provenance, and temporal associations.Filtering validates inputs, execution, and exports; SFT supervision targets planner action tokens, while multi-shot online tasks carry the policy’s committed resources across consecutive requests.Source: §A.3 · p. 15, §A.4 · p. 15

B DATASET STATISTICS AND EXAMPLES

The supplement reports the frozen SFT snapshot, illustrates staged reference access and persistent reuse, and details REEL-Bench inputs and visual-evaluation procedures.

B.1 FROZEN SFT SNAPSHOT

  • The frozen SFT snapshot contains 33,084 unique training trajectory records from 4,142 story identifiers and 3,018 source families; four padding records bring the training set to 33,088 examples.The three construction splits have zero source-family overlap; Table 5 describes accepted shot-level conversations, not benchmark story counts.Source: §B.1 · p. 15, Table 5 · p. 16

B.2 REFERENCE AND REQUEST EXAMPLES

  • Reference examples distinguish uploaded images from source-run frames, and a later-scheduled character upload is unavailable to earlier shots.The staged example shows a scene plate arriving before Shot 4, while the robot reference arriving before Shot 5 cannot condition Shot 4.Source: §B.2 · p. 16, Figure 6 · p. 17

B.3 A RECORDED MEMORY-CREATION AND REUSE EXAMPLE

  • In the recorded rollout, the planner combines retrieved identity evidence, a newly created robot reference, and the prior-shot tail, then reuses the stored reference in Shots 8 and 10.The example has no scheduled user uploads; Figure 7 separately depicts an object reference stored at Shot 3 and retrieved at Shot 10.Source: §B.3 · p. 16, Figure 7 · p. 18

C.1 REEL-BENCH INPUT RELEASE

  • REEL-Bench’s frozen release contains 300 stories, 2,700 shots, 635 upload events, and 609 distinct media files; stories comprise 60 text-only, 210 image-only, and 30 image-plus-voice across eight visual styles.All compared methods receive the same raw requests and scheduled uploads, and each builds memory only from its own preceding outputs; it cannot access another method’s generated images.Source: §C.1 · p. 18

C.2 ST-BENCH METRICS

  • ST-Bench uses five feature-based metrics, including overall and top-10 cross-shot consistency; its top-10 pairs are selected by request similarity.The reimplementation notes that unspecified checkpoint, sampling, and aggregation choices prevent establishing exact numerical equivalence to published results.Source: §C.2 · p. 19

C.3 REEL-BENCH VISUAL JUDGING

  • REEL-Bench judges seven visual dimensions, separating per-shot criteria from story-level consistency, state compliance, and narrative coherence.The story judge sees four chronological sampled frames per shot, so the protocol cannot resolve high-frequency flicker, lip synchronization, or audio quality; human reviewers reported qualitative agreement in inspected examples.Source: Table 7 · p. 20, §C.3 · pp. 20–21

C.4 RELATION TO THE JOYAI-ECHO 1.5 QUANTITATIVE BENCHMARK

The JoyAI-Echo 1.5 report evaluates a separate audiovisual collection, which the authors keep distinct from their measured ST-Bench and REEL-Bench results.

C.4 RELATION TO THE JOYAI-ECHO 1.5 QUANTITATIVE BENCHMARK

  • The JoyAI-Echo 1.5 report evaluates a separate collection of 100 stories and 3,000 shots using visual and audio-related metrics.Its metrics include ViCLIP, Self-CIDS, voice consistency, aesthetic and imaging quality, CLIP text consistency, and speech-content recall.Source: §C.4 · p. 21
  • ST-Bench does not reproduce that audiovisual evaluation, and REEL-Bench supplies no voice-consistency or speech-recall scores.The two benchmarks therefore do not cover the same evaluation dimensions as the reported JoyAI-Echo 1.5 collection.Source: §C.4 · p. 21
  • The original evaluation manifest and complete metric configuration are unavailable, so the paper does not treat the published scores as a third measured benchmark.Source: §C.4 · p. 21

D CORE-100 AGGREGATION

This appendix fixes how Core-100 and generation completion are aggregated, then documents evaluation scope, heuristic baselines, controlled settings, and training configurations.

D CORE-100 AGGREGATION

  • Core-100 combines seven 0–100 dimensions with fixed profiles that weight shot-level scores at 60% and story-level scores at 40%, without renormalizing applicable dimensions.Profiles are selected from frozen requests and available uploads, not method identity or ranking; successful stories receive equal weight in the final mean.Source: D CORE-100 AGGREGATION · p. 21, Table 8 · p. 21

E.1 EVALUATION SCOPE AND CHECKPOINT IDENTITY

  • Main REEL-Bench analyses use 300 stories and 2,700 requested shots; quality scores include only fully completed, validly judged stories, while completion also counts successful shots in incomplete stories.The matched-adapter comparison uses a 30-story, 263-shot subset; its quality comparisons use common successful-story intersections.Source: §E.1 · pp. 22–23, Table 9 · p. 22

E.2 HEURISTIC MEMORY BASELINES

  • Recent Memory fills remaining slots with newest frames, StoryMem-style Memory protects early frames then uses recent ones, and CLIP-based Retrieval ranks historical frames by request similarity.These baselines share the controlled comparison setup, while No Generated Memory retains user uploads but excludes references extracted from generated clips.Source: §E.2 · p. 23

E.3 CONTROLLED GENERATION SETTINGS

  • Per-shot video budgets are capped, but REEL's image-creation, planner, inspection, and adaptation calls add computation; first-attempt seeds are paired within each backend.The controlled comparisons allow up to 64 planner turns, two reference-image creation calls, two video candidates, and one condition-plan revision per shot.Source: §E.3 · p. 24

E.4 TRAINING AND OPTIMIZATION

  • SFT supervises planner actions from interaction histories, then 400 single-shot and 600 multi-shot GRPO updates optimize planner-generated tokens while masking requests, tool outputs, and media.The GRPO configuration uses full-parameter AdamW fine-tuning and samples eight tool-use trajectories per task.Source: §E.4 · pp. 24–25

E.5 ONLINE LEARNING AND ABLATION INTERPRETATION

This appendix reports matched-checkpoint and training-budget comparisons, clarifies the scope of archived uncertainty estimates, and adds illustrative qualitative sequences with explicit limits.

E.5 ONLINE LEARNING AND ABLATION INTERPRETATION

  • Multi-shot GRPO carries policy-produced memory across consecutive requests and adds episode-level memory utility; ablations separately compare training stages and disable retrieval, creation, or review.Retrieval ablation preserves user uploads and preceding-shot context, while unseen-backend transfer changes the adapter and generator with the planner fixed.Source: §E.5 · p. 25
  • 57.69 on Shot and 46.02 on Story: REEL's 400 single-shot plus 600 multi-shot updates outperform single-shot continuation at the same 1,000 updates, using 20,923 rather than 33,871 training video calls.Matched quality uses 213 common successful stories and 1,898 shots; completion is measured over all 2,700 requested shots.Source: Table 12 · p. 26, §E.5 · p. 26

E.6 MATCHED-ADAPTER TRANSFER CONTROL

  • 4.85 and 3.22 points: full REEL improves Story over SFT on Seedance 2.0 and Kling 3.0, respectively, on matched subsets without target-backend training or teacher fallback.Table 13 uses 16 stories (132 shots) for Seedance 2.0 and 18 stories (150 shots) for Kling 3.0; the full-benchmark transfer results are separate.Source: Table 13 · p. 26, §E.6 · p. 26

E.7 SOURCE-RUN UNCERTAINTY

  • Table 14 summarizes an archived JoyAI-Echo 1.5 dashboard run in which all five methods completed and scored the same 300 stories and 2,700 requested shots; it is not the full-benchmark RL interval estimate.Its intervals estimate each method's mean rather than paired differences.Source: §E.7 · p. 27, Table 14 · p. 27

F ADDITIONAL QUALITATIVE COMPARISONS

  • Figure 8 shows midpoint frames from two selected eight-shot cases across four methods; these illustrative examples cannot establish motion smoothness, speech fidelity, or a statistical win rate.The cases use the same requests and upload schedule across methods within each case, but are not a random quality sample.Source: Figure 8 · p. 27, F ADDITIONAL QUALITATIVE COMPARISONS · p. 27

F.1 AN OBSERVED REFERENCE-CREATION FAILURE

This section documents REEL’s reference-creation workflow, from assessing and conditioning on generated evidence to handling failed attempts and completing durable memory. It also describes the prompts, backend adapter, evidence-grounding rules, and blinded shot- and story-level judging used in that process.

F.1 AN OBSERVED REFERENCE-CREATION FAILURE

  • In Shot 2 of The Clockwork Skyfall, a reference already containing suspension lines and decorative stars was misread as five hanging star ornaments, then used as an exact opening frame for a falling-objects event.The error arose before video synthesis and was reinforced by conditioning; the review covered the inspected reference and fixed sampled frames, not full-video motion or audio.Source: §F.1 · p. 28, Figure 9 · p. 28

G COMPLETE SYSTEM PROMPTS

  • The prompt listings reproduce executable construction, collection, SFT, and evaluation system prompts; runtime messages provide current inputs, while only offline construction can inspect the complete source story.Native function schemas define tool arguments, and the listings omit service credentials, private endpoints, and local storage paths.Source: G COMPLETE SYSTEM PROMPTS · p. 29

G.1 REQUEST DIVERSIFICATION AND REVIEW

  • Request diversification prompts generate natural user requests—not planner trajectories—and an independent reviewer checks input validity against the source story and interaction recipes.Requests preserve source events and identities while varying expression, framing, timing, evidence needs, and language; revisions reinterpret the same event rather than advancing story time or alleging failure.Source: §G.1 · pp. 29–30

G.2 INVENTORY AND SIMULATED UPLOAD PREPARATION

  • Upload preparation inventories shot-specific entities and changing states, then designs selected evidence images for one shot rather than treating uploads as an oracle for the whole shot.An independent image check evaluates the final candidate against the request, purpose, description, and preceding references, inspecting pixels rather than trusting generation claims.Source: §G.2 · pp. 30–31

G.3 COLLECTION AND SFT POLICY PROMPTS

  • The collection policy produces both a semantic evidence plan and a complete generation prompt, discovers context through tools, and leaves backend-specific conditioning choices to a separate adapter.The user request remains immutable, and the adapter treats original intent as authoritative when enhancing the draft.Source: §G.3 · pp. 31–32
  • The policy requires viewing and authorizing evidence before use, ranking only useful references, and separating durable identity from transient state and current-scene details.It permits optional image creation for missing useful evidence, requires inspection of decodable candidates, and limits memory updates to durable facts supported by assessed evidence.Source: §G.3 · pp. 32–34
  • Candidate review must distinguish evidence-selection and prompt problems from stochastic or rendering failures, and repair only after an observed failed or uncertain check.When no useful permitted correction or budget remains, the policy requires an honest best-effort decision rather than treating every failure as memory-caused.Source: §G.3 · pp. 32, 35
  • REEL uses the same semantic workflow across video backends, while a separate adapter enhances the draft prompt and maps conditions to each backend’s supported inputs.The policy assigns prompt authoring and memory selection to REEL, but excludes backend endpoints, layouts, seeds, and provider parameters from its choices.Source: §G.3 · p. 33
  • The generation prompt must preserve the immutable user request, while the evidence plan must cover required needs or explicitly report them unfulfilled.Bindings use viewed, authorized representations and semantic roles; exact-opening evidence is appropriate only when the image matches the requested starting pose and composition.Source: §G.3 · pp. 33, 35
  • Repairs must change evidence or prompts in response to observed failures, while resampling preserves conditions and requests a different host-controlled seed.A new candidate ID is not itself a repair, and pending jobs do not authorize duplicate renders.Source: §G.3 · p. 36
  • Durable memory is limited to stable facts supported by eligible inspected evidence or trusted upload provenance, and sessions finish only after required work is resolved.Prompts exclude speculative or transient facts and failed candidates; completion requires closing candidate assessments and staging required memory or explicitly determining that no update is needed.Source: §G.3 · p. 36

G.4 BLINDED VISUAL EVALUATION PROMPTS

  • Shot-level judges independently assess prompt following, reference fidelity, visual quality, and sampled-frame motion using only supplied visual evidence.Reference fidelity uses only images available by that shot, and requested state changes override obsolete reference states; absent expected subjects are failures, not null scores.Source: §G.4 · pp. 36–37
  • Story-level judges assess cross-shot consistency, requested state changes, and narrative coherence from observable clips without inferring hidden events or identities.They distinguish stable identity from changing state, allow intentional cuts and ellipses, and treat a missing expected subject or state as failure rather than null.Source: §G.4 · pp. 37–38

Related papers

Loading 2610.16632v1…