Hugging Face's logo

Datasets:
Alissonerdx
/
h3upscale-4k

Tasks:
Image-to-Image
Video-to-Video
Modalities:
Image
Text
Formats:
text
Languages:
English
Size:
1K - 10K
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
 Community
2

Trained it end to end: h3upscale x4 / x2 LoRAs + a ComfyUI guide node

#2
by sandpies - opened
Discussion
sandpies

Hi Alisson, thanks for preparing this dataset and the latent-guides fork. The card says an end-to-end run hasn't been verified yet, so here is one.

Setup
Your minimax-h3-latent-guides branch, following TRAINING_GUIDE.md: images plus the 111 train videos at num_frames: 5, 4000 steps, rank 64, batch 1, on the pruned int8 Ref2VA base (one RTX PRO 6000, about 1.5 h per run). Two runs, only reference_downscale_factor changed:

  • v1, factor 4: the guide is followed (pose, framing, colour), but at 1/4 size a whole face is only about 75 px wide in the guide, and identity drifts on every seed.
  • v2, factor 2: on stills it stays very close to the source (768x1152 from a 384x576 guide; 1920x1088 from 960x544).

Video findings with v2

  • Generating from noise with the guide flickers: brightness snaps back every 17 frames, on each group's single-frame key token. My reading: 5-frame clips only teach the first two latent steps of a 17-frame group.
  • Refining an existing clip (start from the clip, light denoise, guide = the clip at half size) is stable: no flicker, detail added, motion and mouths kept.
  • In a tiled latent upscale (H3 latent upscaler, then one re-draw step at denoise 0.1 with the guide cut per tile), the result beat the native render in a blind side-by-side. On the bare DiT it looks cleaner than with other LoRAs stacked under it.

Suggestion
All 111 train clips have at least 39 frames (they are about 2 s each), so num_frames: 39 would use every clip and cover two full 17-frame groups. That is probably what generation from noise needs.

ComfyUI
Your guide notes that integrations need the reduced-resolution packed layout. I implemented it as a small PackedLayout patch for keyframes marked with a stride, and checked it row for row against your build_packed_sequence (row counts, positions, row classes).

Code and the v2 LoRA:

v1 isn't published; happy to share it if it's useful. Credit to you for the dataset, the fork and the guide layout is in the repo's notices.
Alissonerdx
 Owner

Hi Alisson, thanks for preparing this dataset and the latent-guides fork. The card says an end-to-end run hasn't been verified yet, so here is one.

Setup
Your minimax-h3-latent-guides branch, following TRAINING_GUIDE.md: images plus the 111 train videos at num_frames: 5, 4000 steps, rank 64, batch 1, on the pruned int8 Ref2VA base (one RTX PRO 6000, about 1.5 h per run). Two runs, only reference_downscale_factor changed:

  • v1, factor 4: the guide is followed (pose, framing, colour), but at 1/4 size a whole face is only about 75 px wide in the guide, and identity drifts on every seed.
  • v2, factor 2: on stills it stays very close to the source (768x1152 from a 384x576 guide; 1920x1088 from 960x544).

Video findings with v2

  • Generating from noise with the guide flickers: brightness snaps back every 17 frames, on each group's single-frame key token. My reading: 5-frame clips only teach the first two latent steps of a 17-frame group.
  • Refining an existing clip (start from the clip, light denoise, guide = the clip at half size) is stable: no flicker, detail added, motion and mouths kept.
  • In a tiled latent upscale (H3 latent upscaler, then one re-draw step at denoise 0.1 with the guide cut per tile), the result beat the native render in a blind side-by-side. On the bare DiT it looks cleaner than with other LoRAs stacked under it.

Suggestion
All 111 train clips have at least 39 frames (they are about 2 s each), so num_frames: 39 would use every clip and cover two full 17-frame groups. That is probably what generation from noise needs.

ComfyUI
Your guide notes that integrations need the reduced-resolution packed layout. I implemented it as a small PackedLayout patch for keyframes marked with a stride, and checked it row for row against your build_packed_sequence (row counts, positions, row classes).

Code and the v2 LoRA:

v1 isn't published; happy to share it if it's useful. Credit to you for the dataset, the fork and the guide layout is in the repo's notices.

Hahaha, that's cool! I'm glad you tested it. I actually ended up making that repo public, even though it was supposed to be private; I have so many datasets that I hit my storage limit, so I decided to just leave it public. It's great that it helped, I'll test your model later.

Sign up or log in to comment