Local finishing pass for final Renditions #15

New issue

Closed

opened 2026-10-06 03:55:55 +00:00 by crowmaki · 1 comment

crowmaki commented

2026-10-06 03:55:55 +00:00

Owner

Part of #1

Question

Which local video upscaler best replaces the hosted H3-Regenerate-2K on a single 32 GB GPU? Compare it with the H3 latent upscaler (minimax_h3_latent_upscaler_3d) used in the seed-hunter workflow. What frame-interpolation policy (RIFE, target fps) makes sense for final Renditions delivered to an editor?

Part of #1 ## Question Which local video upscaler best replaces the hosted H3-Regenerate-2K on a single 32 GB GPU? Compare it with the H3 latent upscaler (`minimax_h3_latent_upscaler_3d`) used in the seed-hunter workflow. What frame-interpolation policy (RIFE, target fps) makes sense for final Renditions delivered to an editor?

crowmaki added the

wayfinder:research

label

2026-10-06 03:55:55 +00:00

crowmaki referenced this issue

2026-10-06 03:56:04 +00:00

Shipped workflow roster and slot contracts #17

crowmaki referenced this issue

2026-10-06 03:56:08 +00:00

Wayfinder: AI Studio end-to-end slice spec #1

crowmaki self-assigned this

2026-10-06 03:56:44 +00:00

crowmaki referenced this issue from a commit

2026-10-06 03:59:41 +00:00

research: local finishing pass for final Renditions (#15)

crowmaki commented

2026-10-06 03:59:53 +00:00

Author

Owner

Resolution

Recommendation

Primary finishing path: the H3 latent upscaler 3D, then an H3 refine pass. This is what the seed-hunter workflow already does. It is the closest local stand-in for H3-Regenerate-2K, which is hosted-only and has no release date. H3 itself regenerates the detail, so the look and the motion over time stay native to the model. This is also exactly how Promotion works: it reuses the draft latent, with no round trip through the ~5B VAE. The catch, in the authors' words: it "saves time, not VRAM". The refine pass runs the 33B model at the target size. Peak VRAM and runtime for 1080p/2K on the 5090 are not published and must be measured.
Fallback, or the step after the refine for 2K/4K delivery: SeedVR2 (7B fp16 fits 32 GB; the 1.4B distill is faster). It is the strongest pixel-space option, at about 312 s for 1080p on a 5090 per the ComfyUI docs. Use small scale factors: its authors warn it over-generates detail on light-degradation 720p AIGC input.
Don't use FlashVSR for finals. It is tuned for 4x only, its 2x output is poor, and the ComfyUI ports lack its sparse attention.
Frame rate: finals at native 24 fps, no RIFE. 24 to 60 invents 3 of every 5 frames at a non-integer cadence, and an editor working at 24 fps either drops those frames or gets judder and a video look. Keep RIFE as an opt-in derived file (e.g. 2x for an intended slow-mo), never as the final Rendition. Wire the final VHS_VideoCombine from the refine output at 24 fps, not from the RIFE node.

Context: branch research/finishing-pass, file research/finishing-pass.md. It has the option table, sources, the workflow inspection (both nodes are currently bypassed) and the open items to measure.

## Resolution **Recommendation** - **Primary finishing path: the H3 latent upscaler 3D, then an H3 refine pass.** This is what the seed-hunter workflow already does. It is the closest local stand-in for H3-Regenerate-2K, which is hosted-only and has no release date. H3 itself regenerates the detail, so the look and the motion over time stay native to the model. This is also exactly how **Promotion** works: it reuses the draft latent, with no round trip through the ~5B VAE. The catch, in the authors' words: it "saves time, not VRAM". The refine pass runs the 33B model at the target size. Peak VRAM and runtime for 1080p/2K on the 5090 are not published and must be measured. - **Fallback, or the step after the refine for 2K/4K delivery: SeedVR2** (7B fp16 fits 32 GB; the 1.4B distill is faster). It is the strongest pixel-space option, at about 312 s for 1080p on a 5090 per the ComfyUI docs. Use small scale factors: its authors warn it over-generates detail on light-degradation 720p AIGC input. - **Don't use FlashVSR for finals.** It is tuned for 4x only, its 2x output is poor, and the ComfyUI ports lack its sparse attention. - **Frame rate: finals at native 24 fps, no RIFE.** 24 to 60 invents 3 of every 5 frames at a non-integer cadence, and an editor working at 24 fps either drops those frames or gets judder and a video look. Keep RIFE as an opt-in derived file (e.g. 2x for an intended slow-mo), never as the final Rendition. Wire the final `VHS_VideoCombine` from the refine output at 24 fps, not from the RIFE node. **Context:** branch `research/finishing-pass`, file `research/finishing-pass.md`. It has the option table, sources, the workflow inspection (both nodes are currently bypassed) and the open items to measure.

crowmaki closed this issue

2026-10-06 03:59:54 +00:00

crowmaki referenced this issue

2026-10-06 04:01:22 +00:00

Verify latent round trip and refine-pass cost #20

crowmaki referenced this issue

2026-10-07 07:24:15 +00:00

Spec: AI Studio end-to-end slice #24

crowmaki referenced this issue

2026-10-07 17:39:12 +00:00

Shipped workflow templates and manifests (#33) #79