Hyperflow retention-analysis comparison

Open the interactive presentation for paired playback and downloadable comparison videos.

This Hugging Face Static Space runs entirely in the browser and does not require GPU hosting.

All twenty original clips and ten comparisons are complete. See report.md for the findings. Original renders, viewing copies, side-by-side videos, reference images, exact prompts and sampled frames are included. SHA256SUMS records the presentation files.

Design

Twenty independently rendered clips form ten fixed-seed A/B pairs. All clips use five publicly sourced reference images, a long English Ref2VA prompt, full Hyperflow with its trained eight-step schedule, Euler, CFG 1, and the same attention backend. There is no SelfLift/upscale stage or diffusion-cache node. The user's production workflow is not used.

Pairs A B Question
01–05 Fully-preserved retention-analysis block Entire retention-analysis block omitted Does that redundant block help in complex multi-reference scenes?
06–10 Appropriate partially-preserved, attribute-transfer, or weak-reference markers fully_preserved substituted for those markers only Does the marker change the result when the explicit prose is identical?

The first group tests the practical value of the whole block, including its extra prose. It does not isolate the mode token. The second group does isolate marker text; B deliberately contradicts the requested edits so that the influence of that text can be observed.

A/B rendering order alternates between scenes. Seeds differ between scenes but are identical within each pair. Automated checks verified the planned prompt differences and the executed inputs. Input hashes, source URLs, seeds and exact settings are in manifest.json; the original MP4 metadata embeds the executed ComfyUI graph. Exact A/B prompts are provided in each pair's folder.

How to assess

Compare corresponding timestamps and watch the complete clip. Evaluate these separately:

  1. Reference fidelity: facial appearance, hair, clothing, camera silhouette, puppy appearance, and distinctive café fixtures.
  2. Instruction adherence: specified actions, positions, camera movement, and—where requested—clothing/color/lighting/environment changes.
  3. Temporal behavior: consistency through movement and occlusion, object permanence, contact/hand errors, and unwanted cuts.
  4. General visual quality: texture, sharpness, distortion, and exposure. More visual change is not necessarily an improvement.

Do not assume a mode is ignored because both clips achieve the same requested outcome. Conversely, do not assume a mode is useful merely because images differ. One seed per scenario, five scenarios per claim, one shared reference set, and no same-prompt repeat control make this an exploratory comparison, not statistical proof or a finding about the model's internal token handling.

Files

Each pairs/<scene>/ folder contains native originals with embedded execution metadata, exact 10-second A/B viewing copies, a side-by-side MP4 with A's audio, timestamped full-resolution frames, a paired contact sheet and both prompts. Native outputs contain 243 frames at 24 fps (10.125 seconds); viewing copies are trimmed to 240 frames without changing speed.

The HTML gallery lets you switch between A and B audio. Audio has not been assessed in this study; image contact sheets alone cannot establish audio quality or lip-sync accuracy.

Sources

The tested claims were supplied in screenshots of Ryko's H3 Chatter discussion: removing a fully-preserved retention block reportedly produced similar results, and explicit natural-language instructions reportedly overrode preservation-mode markers. The full Discord conversation was not retrieved.

Reference images are public Unsplash image URLs recorded individually in manifest.json. H3 prompt format follows the official Ref2VA guide; Hyperflow setup follows its official README.