dagloop5 commited on
Commit
306f6c8
·
verified ·
1 Parent(s): cea8f29

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +0 -66
app.py CHANGED
@@ -762,72 +762,6 @@ with gr.Blocks(title="PlagueKind · MiniMax-H3") as demo:
762
  lora_b_strength,
763
  ]
764
 
765
- # Two blocks rather than one with empty keyframe cells: the positional order below matches `generate`'s
766
- # signature, so each block fills a prefix of it and the defaults cover the rest.
767
- gr.Examples(
768
- label="Text to video",
769
- examples=[
770
- ["A red fox trotting through a snowy pine forest at dawn, snow crunching underfoot", DEFAULT_CANVAS],
771
- ["A busy night market, neon signs reflecting in puddles, sizzling street food", "544x960 · 9:16 fast"],
772
- ["A cellist playing a slow melody in an empty concert hall", "544x544 · 1:1 fast"],
773
- ],
774
- inputs=[prompt, canvas],
775
- outputs=[video, report],
776
- fn=generate,
777
- cache_examples=True,
778
- cache_mode="lazy",
779
- )
780
- gr.Examples(
781
- label="First and last frame",
782
- examples=[
783
- [
784
- "A slow seamless camera move from the first view to the last, wind in the pines",
785
- "1152x640 · 16:9",
786
- "examples/first.png",
787
- "examples/last.png",
788
- ],
789
- [
790
- "The fox looks around, then trots deeper into the forest",
791
- "1152x640 · 16:9",
792
- "examples/first.png",
793
- "examples/first.png",
794
- ],
795
- ],
796
- inputs=[prompt, canvas, first_frame, last_frame],
797
- outputs=[video, report],
798
- fn=generate,
799
- cache_examples=True,
800
- cache_mode="lazy",
801
- )
802
-
803
- with gr.Accordion("What this workflow changes, and what it cannot", open=False):
804
- gr.Markdown(
805
- """
806
- `Plaguekind/Minimax-H3` ships no weights — it is a ComfyUI graph over `Comfy-Org/MiniMax-H3`. Its nodes map onto
807
- this Space as:
808
- | ComfyUI node | widget | here |
809
- |---|---|---|
810
- | `KSamplerSelect` | `euler` | MiniMax-H3's only sampler; the checkpoint is CFG-distilled, so one forward per step and no negative prompt |
811
- | `BasicScheduler` | `linear_quadratic`, 15 steps | **Sigma schedule** / **Steps** |
812
- | `MiniMaxH3ImageToVideo` | prompt, first/last frame | **Prompt** / **First frame** / **Last frame** |
813
- | `UnifiedResizeImageMask` ("Target Dimension") | 1344x768 | **Target dimension** |
814
- | `ImageSharpenKJ` | `rcas`, 0.3 | **RCAS sharpening** |
815
- | `FrameInterpolate` + `film_net_fp16` | multiplier 2 | **FILM frame interpolation** |
816
- | `CreateVideo` | fps `24 * 2` | 48 fps output |
817
- | `RTXVideoSuperResolution` | 2x `ULTRA` | **not reproduced** |
818
- | `PathchSageAttentionKJ` | `sageattn_qk_int8_pv_fp8_cuda++` | cuDNN fused attention |
819
- Two deliberate deviations. `RTXVideoSuperResolution` is NVIDIA's NGX super-resolution, shipped as a driver-level
820
- Windows/RTX component with no Linux Python path, so the 2x upscale is missing — pick a larger **Target dimension**
821
- instead of upscaling a small one. And SageAttention's `qk_int8_pv_fp8_cuda++` kernel is not built for this pool's
822
- sm120 cards, so attention runs cuDNN's fused kernel, which is the fastest available here and is numerically the
823
- faithful one (SageAttention is a quantized approximation).
824
- One upgrade: the workflow loads `minimax_h3_fl2va_pruned_int8_convrot.safetensors` and a
825
- `qwen3vl_32b_..._int8_convrot` text encoder because that is what fits a consumer card. This Space runs both
826
- **unquantized bfloat16**, off `MiniMaxAI/MiniMax-H3`, with the 62 GiB text encoder in a
827
- [second Space](https://huggingface.co/spaces/multimodalart/qwen3vl-conditioner).
828
- """
829
- )
830
-
831
  run.click(generate, controls, [video, report], api_name="generate")
832
 
833
 
 
762
  lora_b_strength,
763
  ]
764
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
765
  run.click(generate, controls, [video, report], api_name="generate")
766
 
767