MiniMax H3 (local): text prompt has zero effect on output — same seed + different prompts produce pixel-identical videos

# GitHub Issue Draft: MiniMax H3 — text conditioning has zero effect on output (same seed + different prompts → bit-identical videos)

## Summary

When running MiniMax H3 locally via the built-in `MiniMaxH3ImageToVideo` node, the text prompt has **no effect whatsoever on the generated output**. Two completely different prompts with the same seed produce **pixel-identical videos**. The model appears to run in an unconditional mode: sampling works (different seeds → different outputs), but conditioning never influences generation. The decoded output is always a structured mosaic of 16×16 colored blocks (green/magenta pattern), which looks like the VAE faithfully decoding latents that have patch-aligned structure but no semantic content.

## Environment

- ComfyUI: v0.33.x (tested on `72865f4f` v0.33.1 and latest master `7dde5617`)
- OS: Windows 11, GPU: NVIDIA RTX 4060 Ti 16GB, torch 2.12.1+cu130
- comfy-aimdo 0.4.13, comfy-kitchen 0.2.31
- Models:
- Diffusion: `minimax_h3_fl2va_pruned_int8_convrot.safetensors` (19.53 GB, from Comfy-Org/MiniMax-H3)
- Text encoder: `qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors` (14.61 GB)
- VAEs: `minimax_h3_video_vae_fp16.safetensors`, `minimax_h3_audio_vae_fp32.safetensors`
- LoRA (optional): `minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors`
- Submitted via `POST /prompt` (API format). Also reproduced via UI-built graph.

## Minimal reproduction

Standard t2v chain:

```
UNETLoader → LoraLoader(Turbo, optional) → MiniMaxH3SigmaShift(12/3)
CLIPLoader(type=minimax) ──────────────────────────────┐
VAELoader(video) ─────────────────────────────────────┐│
↓↓
MiniMaxH3ImageToVideo(prompt=X, 832x480, length=22)
│ CONDITIONING │ LATENT
↓ ↓
BasicGuider SamplerCustomAdvanced
└────── VAEDecode → CreateVideo → SaveVideo
```

Run A: prompt = "A cute orange tabby cat sitting on a windowsill..." , seed = 111
Run B: prompt = "A vast dark ocean at night with huge waves crashing...", seed = 111

**Result: outputs are pixel-identical (mean abs diff = 0.0000 across all frames).**

Control: same prompt, seed 111 vs seed 999 → outputs differ (mean abs diff ≈ 116), so the sampler/noise path works.

This holds with and without the Turbo LoRA, at length 22 and length 124 (in-training-range), at 864×480 and 832×480.

## What I verified (so the report narrows the search space)

I instrumented the graph with a passthrough diagnostic node and checked every stage:

1. **Server receives correct inputs** — `/history` shows both graphs carried their distinct `prompt` strings into node 6 (`MiniMaxH3ImageToVideo`). Not an API-format field-dropping issue.
2. **Text encoder works** — loading the TE directly via `comfy.sd.load_clip(CLIPType.MINIMAX)` outside the server, "cat…" encodes to `[1, 26, 5120]`, "ocean…" to `[1, 23, 5120]`. Different token counts, different embeddings. Tokenizer fine.
3. **Conditioning content is correct inside the graph** — a passthrough node logged, for each prompt:
```
emb shape=[1, 26, 5120] mean=0.153 std=42.85
meta keys: ['minimax_token_tags', 'pooled_output']
minimax_token_tags values: {1: 26} ← all text-modality tags, correct
```
So `clip.tokenize` → `encode_from_tokens_scheduled` produces sane, distinct conditionings that reach the guider.
4. **VAE is fine** — decoding the *empty* AV latent from `MiniMaxH3ImageToVideo` (skipping sampling entirely) yields a uniform dark-gray image, as expected for zeros. Decode path healthy.
5. **Not the comfy-kitchen CUDA INT8 kernels** — force-disabled the CUDA backend (`ck.registry.disable("cuda")`, forcing eager/emulated dequant) and re-ran: same-prompt output differs from the CUDA-backend run by only 0.0096 (float noise), different-prompt still 0.0000. Kernels exonerated.
6. **Not aimdo/DynamicVRAM paging corruption** — identical blocky output with DynamicVRAM enabled and with `--disable-dynamic-vram` (legacy lowvram loader).
7. **Not the LoRA** — reproduces with `strength_clip=0`, `strength_model=0`, and with LoRA node removed entirely.
8. **Not guider choice** — `BasicGuider` and `CFGGuider(cfg=4, negative=)` produce identical results.
9. **No runtime errors** — runs report success; no NaN warnings; models load fully ("208 patches attached" with Turbo).

## Interpretation

Everything upstream of the DiT is correct and distinct per prompt, yet the sampled latent does not depend on the prompt at all and has deterministic patch-aligned block structure. This points at the conditioning → DiT handoff inside the H3 forward path: e.g. the packed sequence built in `MiniMaxH3.extra_conds` / `PackedLayout` / `_forward` not actually admitting the `c_crossattn` rows into the attention computation on this setup, leaving the model to denoise unconditionally.

Suspect area: interaction between the `int8_convrot` weight format of `minimax_h3_fl2va_pruned_int8_convrot.safetensors` and the H3 packed-layout forward in ComfyUI 0.33.x (possibly the `preprocess_text_embeds` / `condition_proj` path, or cross-attn rows being dropped when building the packed layout).

## Requested

- Confirmation whether others can reproduce with `minimax_h3_fl2va_pruned_int8_convrot` on NVIDIA (esp. RTX 40-series), API or UI submission.
- Any known-good combo (ComfyUI version / backend flags / alternative quantized checkpoint) where text conditioning demonstrably changes the output with this checkpoint.

Happy to provide full workflow JSONs, logs, and output frames.

Trace the H3 conditioning handoff through MiniMaxH3.extra_conds, PackedLayout, and _forward, focusing on preprocess_text_embeds and condition_proj. Compare the packed inputs and attention path for the two prompts, then reproduce with the listed checkpoint and workflow. Done means identical-seed runs with distinct prompts produce different sampled latents or decoded videos while the existing seed control remains valid.

Written by the indexing model from the issue text.