MiniMax H3 GGUF MiniStack

What's in the box

📂 ComfyUI/models/
├── 📂 diffusion_models/
│   └── MiniMax-H3-FL2VA-pruned-Q4_K_M.gguf       11.38 GB  pruned AdaLN DiT, Q4_K_M
├── 📂 text_encoders/
│   ├── qwen3vl-4b-h3student-Q4_K_M.gguf           2.50 GB  distilled TE, Q4_K_M (default)
│   ├── qwen3vl-4b-h3student-BF16.gguf             8.05 GB  distilled TE, BF16 (optional)
│   ├── te_adapter_v1.safetensors                    120 MB  student→H3 5120-d MLP adapter (required)
│   └── h3_tokenizer/                                  11 MB  H3 tokenizer (required)
└── 📂 vae/
    └── minimax_h3_video_vae_fp16.safetensors       4.85 GB  stock video VAE, fp16

Documented changes vs. stock

1. Pruned DiT (FL2VA). The 33B full-width DiT is replaced with the AdaLN-curve form from Comfy-Org/MiniMax-H3 (minimax_h3_fl2va_pruned_bf16): AdaLN 13B→44M params via the adaln_t_table [1025,8] + per-block [96768,8] projections. Effective ≈ 18.7B params.
2. Q4_K_M quantization. Converted with a two-pass streaming converter and quantized with a ComfyUI-GGUF-patched llama.cpp build (arch "wan"; stock llama-quantize rejects this arch). 11.38 GB vs 14.5 GiB for the full Q3_K_M. GGUF: general.architecture=wan, 532 tensors, 53 fp32/fp16 hiprec islands (adaln*, t_embed, norm, patch_proj) kept as-is.
3. Distilled student text encoder. The 32B Qwen3-VL-32B teacher is replaced with a Qwen3-VL-4B student plus a trained MLP adapter (2560→4096→5120, fp32) that maps student language hidden states into H3's 5120-d multimodal token slots. Distilled from 2,500 teacher dumps; val_cos 0.9310 (full-corpus v4 adapter). Q4_K_M GGUF is numerically near-identical to the BF16 student (>0.99 cosine); the gap to the teacher is adapter fidelity, not quant noise.
4. Audio VAE dropped. The minimax_h3_audio_vae_fp32 component is removed from the workflow. Output videos are silent. The stock minimax_h3_video_vae_fp16 is unchanged and included.
5. Video VAE unchanged. Stock fp16 video VAE; no pruning or requantization was applied.

Requirements

• ComfyUI with MiniMax H3 support (Comfy-Org/ComfyUI PR #15224 / current main) and the ComfyUI-GGUF custom node.
• For the student TE you need the h3_small_te custom node (H3SmallTELoader / H3SmallTextEncoder) — it loads the GGUF via ComfyUI-GGUF's gguf_clip_loader into comfy-native Llama2_(Qwen3VL_4BConfig) with GGMLOps, keeping the H3 HF tokenizer + Identity final-norm + the adapter. The adapter and h3_tokenizer/ are required — without them the student cannot encode.
• DiT and video VAE are loaded with the standard UnetLoaderGGUF / VAELoader nodes.

Recommended settings (640×480 previz, Arc A770 16 GB)

• Sampling: 8 steps with the turbo LoRA (larryvrh/MiniMax-H3-Turbo-Lora), --reserve-vram 3. Note the turbo LoRA's adaln_proj targets dangle on the pruned checkpoint — apply it to the full checkpoint if you need those.
• Length: snap to 5 mod 17 frames (17k+5 grid). Example: length = max(5, round(seconds*24)) + (5 - (max(5, round(seconds*24)) % 17)) % 17.
• Default: 640×480 / 124 frames / 24 fps / 8 steps.

Limitations

• No audio. The audio VAE is dropped from this stack — generated videos are silent and there is no voice/SFX/music output.
• Student TE is text-only. The 4B student adapter was distilled on pure-text teacher dumps; it has no vision path. For I2V / FL2V keyframe identity (which needs the teacher's vision tower + deepstack features), use the teacher 32B TE (qwen3vl-32B-MiniMax-H3-*); the student cannot consume image tokens.
• Adapter fidelity ceiling. val_cos ≈ 0.93 vs the teacher. Shot structure and cuts are preserved, but cut timing can drift ±2 s and fine prompt specifics (small props, signage details) can wash out. Timestamp-magnitude tuning (v3/v4) plateaued; this is accepted as a known tradeoff.
• Turbo LoRA compatibility. The preview turbo LoRA targets full-width adaln_proj weights that do not exist in the pruned curve-form checkpoint; using both together silently skips those targets.
• GGUF arch. The DiT is archived as "wan" for quantizer compatibility; use a ComfyUI-GGUF build that can load arch-wan GGUFs. The student GGUF is arch qwen3vl.
• Frames-vs-seconds prompting. H3 is trained on mm:ss.mmm timestamps; the 4× temporal VAE grid only quantizes where cuts can land — frame-based timestamps are off-distribution.
• Not for 2K. Local gen is 640×480 previz; 768P/2K regeneration requires the official MiniMax cloud API.

Usage

1. Unzip files into the ComfyUI models/ folders as shown above.
2. For T2V (text only): UnetLoaderGGUF → pruned Q4 GGUF; H3SmallTELoader → student Q4 GGUF + adapter + tokenizer dir; VAELoader → video VAE.
3. For I2V / FL2V: swap the student TE for the teacher 32B TE (vision path); everything else stays.
4. Encode with MiniMaxH3ImageToVideo, sample with SamplerCustomAdvanced (res_multistep / simple, 25 steps, or 8 with turbo LoRA).

Eval notes (summary)

• Baseline (full Q3_K_M, 25 steps): 620 s wall @ 640×480/124f.
• Pruned Q4_K_M (25 steps): 610 s wall, output on-par with baseline (same composition/cut).
• Turbo 4/8-step: ~160 s / ~250 s wall; 8-step recommended.
• FL2V smoke (pruned Q4 + teacher TE): 295.5 s wall; keyframe lock first/last pixel-cos 0.9905/0.9987.
• Student vs teacher numeric A/B: mean cosine 0.847 (mid adapter) → 0.9310 (full-corpus v4). Video A/B: structure and cuts preserved; fine details degrade consistent with ~0.93 cos.

License