[docs] simplify FastH3 model card

README.md

@@ -25,83 +25,40 @@ tags:
 
 # FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree
 
-FastH3 4-Step Preview v1 checkpoint from FastVideo. This variant is the
-`dmd2_sp4_fsdp32_v12_datafree_mixed_dense_fa4` data-free DMD2 MiniMax-H3
-student at training step 1000. It generates synchronized video and audio with
-four transformer forwards and uses dense FlashAttention.
+A dense-attention FastH3 Preview v1 ablation from
+[FastVideo](https://github.com/hao-ai-lab/FastVideo). It generates synchronized
+video and audio from text with four transformer forwards. This step-1000 model
+was trained with data-free DMD2.
 
-See the complete [FastH3 4-Step Preview v1
-collection](https://huggingface.co/collections/FastVideo/fastvideo-fasth3).
+[Blog](https://haoailab.com/blogs/fasth3-preview/) ·
+[Matching LoRA](https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA/tree/main/dense-datafree) ·
+[Recommended checkpoint](https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree) ·
+[FastH3 collection](https://huggingface.co/collections/FastVideo/fastvideo-fasth3)
 
-## Checkpoint identity
-
-- Training run: `20260826T014507Z-v12-datafree-dense-fa4-r300-rack2`
-- Lineage parent: `20260825T010407Z-v12-datafree-dense-fa4`
-- W&B: [bdpc9m5x](https://wandb.ai/wlsaidhi/h3-dmd2-vsa/runs/bdpc9m5x)
-- Step: 1000
-- Transformer content SHA-256: `3f56129acb9749f2ae4e52bab176c09406bf027d68c6f020cf5ff99e651b1609`
-- Checkpoint metadata SHA-256: `a1ae1a4ec117b5bc79f61da4ee5d4003694f27dc6dfe55e9166b59d2b387e6b9`
-- FastVideo execution commit: `aa6e6c1c1e15f879011a47996c7d6bc8381eaf89`
-- Data commit: `b92be606a439c67006220459df236d18c48fd38c3e11f4771d21fe0d078a3371`
-
-`checkpoint_content.json` records every indexed transformer shard and digest.
-`checkpoint_metadata.json` preserves the source training configuration.
-
-## Contents
-
-This is a self-contained Diffusers modular-pipeline package. The distilled
-bf16 student is under `transformer/`; the text encoder, tokenizer, processor,
-video VAE, audio VAE, and schedulers are unmodified MiniMax-H3 components.
-
-The `transformer_ref` component is not bundled because this student targets
-T2AV. Its modular-index entry remains on `MiniMaxAI/MiniMax-H3`.
-
-## Exact FastVideo inference contract
-
-The student was trained on the explicit ladder `[999, 749, 500, 250]`. These
-are four transformer forwards; do not replace the ladder with an arbitrary
-four-point scheduler grid.
+## Run with FastVideo
 
 ```bash
-export FASTVIDEO_DMD_DENOISING_STEPS=999,749,500,250
-export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
+git clone https://github.com/hao-ai-lab/FastVideo.git
+cd FastVideo
+uv venv --python 3.12
+source .venv/bin/activate
+UV_TORCH_BACKEND=cu130 uv pip install -e ".[fasth3]"
+```
 
+```bash
+FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN \
 python examples/inference/basic/basic_minimax_h3_t2v.py \
   --model-path FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree \
-  --prompt '<MiniMax-H3 T2AV prompt>' \
-  --steps 5 \
-  --num-gpus 4
+  --prompt "your prompt" \
+  --steps 5
 ```
 
-The five scheduler grid points execute the four trained jumps. Use guidance
-scale 1.0 and dense FlashAttention (the FA4 route on the pinned B200 stack).
-Compilation is an optional runtime optimization; first validate output parity
-on the pinned FastVideo commit.
-
-## Training summary
-
-- Method: data-free DMD2 with carried backward-simulation ODE rollouts.
-  Training consumes text conditioning and student-generated latents rather
-  than target video latents.
-- Student grid: `[999, 749, 500, 250]`.
-- Student attention: dense FlashAttention (the FA4 route on the pinned B200
-  stack).
-- Global batch size: 64.
-- Student and critic learning rate: `2e-06` and
-  `2e-06`.
-- Precision: fp32 training state and bf16 inference export.
-- Training topology: 32 B200 GPUs, SP=4,
-  HSDP replicate=1 / shard=32.
-- Continuous score-clock shift: `2.4`; warp max: `0.999`.
-
-## Limitations
+Five scheduler points execute the trained four transformer forwards. Adjust
+`--num-gpus` for your setup.
 
-- Intermediate checkpoint at step 1000 of a planned 4000-step run.
-- T2AV only; the reference-conditioned transformer was not distilled here.
-- Quality remains below base MiniMax-H3 on difficult motion, fine structure,
-  and some audio details.
-- Sampling outside the trained ladder or attention geometry is
-  off-distribution.
-- Inherits the MiniMax H3 Community License and base-model limitations.
+## Scope
 
-Indexed student: 638 bf16 tensors / 66,245,985,792 logical bytes.
+This preview supports text-to-audio-video generation. FL2VA and Ref2VA were
+not distilled. Difficult motion, fine detail, and some audio may remain below
+the base MiniMax H3 model. This checkpoint inherits the
+[MiniMax H3 Community License](LICENSE).