Load Comfy-Org's pruned int8_convrot diffusion model (H3_DIFFUSION_MODEL switch)

README.md

@@ -70,6 +70,7 @@ editing `app.py`.
 | `H3_PRELOAD_MODELS` | `1` | run the four loader nodes at startup so the GPU worker inherits the weights from the parent's RAM instead of reading ~34 GB from disk inside every job's GPU window |
 | `H3_GPU_RESIDENT` | `0` | also ask ComfyUI's model management to place the preloaded patchers on `cuda` at module level, so ZeroGPU restores them on allocation (the way the Wan Space's diffusers pipeline is loaded). Opt-in until it has been seen to work under the CUDA emulation; a failure falls back to the CPU preload |
 | `H3_SAGE_ATTENTION` | `0` | enable the SageAttention patch (INT8 quantised attention — faster, near-lossless, not bit-exact). Needs the `sageattention` wheel to import; the Space log says whether it did |
+| `H3_DIFFUSION_MODEL` | `int8` | which conversion of the diffusion model to load — see *Models* |
 | `MIN_GPU_SECONDS` / `MAX_GPU_SECONDS` | `60` / `1800` | clamp on the declared duration |
 
 Keyframes are scaled to the canvas with `lanczos` (upstream used `nearest-exact`,
@@ -85,7 +86,7 @@ When `job_id` is set and Space secret `HF_TOKEN` has write access, the finished
 
 | role | file | repo |
 |---|---|---|
-| diffusion model | `minimax_h3_fl2va_pruned_nvfp4.safetensors` (12.5 GB) | `lilcheaty/MiniMax-H3-NVFP4` |
+| diffusion model | `minimax_h3_fl2va_pruned_int8_convrot.safetensors` (21 GB) — Comfy-Org's recommended I2V file; `H3_DIFFUSION_MODEL=fp8` → `…pruned_fp8_scaled` (21 GB), `nvfp4` → lilcheaty's `…pruned_nvfp4` (12.5 GB, the original choice: it lost prompt adherence against Wavespeed's open-weights H3 on 2026-10-16) | `Comfy-Org/MiniMax-H3` |
 | text encoder | `qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors` (15.7 GB) | `Comfy-Org/MiniMax-H3` |
 | video VAE | `minimax_h3_video_vae_fp16.safetensors` (5.2 GB) | `Comfy-Org/MiniMax-H3` |
 | audio VAE | `minimax_h3_audio_vae_fp32.safetensors` (0.6 GB) | `Comfy-Org/MiniMax-H3` |

app.py

@@ -37,10 +37,28 @@ COMFY_OUTPUT = COMFY / "output"
 # ComfyUI master on 2026-09-12; requirements.txt mirrors this commit's own pins.
 COMFYUI_COMMIT = os.environ.get("COMFYUI_COMMIT", "c75d8c966c29cb0392259af791f43373315b72db")
 
+# Diffusion model conversion. The community NVFP4 file the Space started on lost prompt
+# adherence against Wavespeed's open-weights H3 (same prompt, image and settings,
+# 2026-10-16); Comfy-Org's own pruned int8_convrot is the file its I2V template
+# recommends. "pruned" drops the precomputable AdaLN branches, not quality. The model
+# card says to prefer int8_convrot where torch has cu130 and to use fp8_scaled only
+# when it can't — switch with the Space variable H3_DIFFUSION_MODEL (int8 | fp8 | nvfp4).
+DIFFUSION_MODELS = {
+    "int8": ("minimax_h3_fl2va_pruned_int8_convrot.safetensors",
+             "Comfy-Org/MiniMax-H3", "diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors"),
+    "fp8": ("minimax_h3_fl2va_pruned_fp8_scaled.safetensors",
+            "Comfy-Org/MiniMax-H3", "diffusion_models/minimax_h3_fl2va_pruned_fp8_scaled.safetensors"),
+    "nvfp4": ("minimax_h3_fl2va_pruned_nvfp4.safetensors",
+              "lilcheaty/MiniMax-H3-NVFP4", "minimax_h3_fl2va_pruned_nvfp4.safetensors"),
+}
+DIFFUSION_MODEL = os.environ.get("H3_DIFFUSION_MODEL", "int8").strip().lower()
+if DIFFUSION_MODEL not in DIFFUSION_MODELS:
+    print(f"[setup] unknown H3_DIFFUSION_MODEL={DIFFUSION_MODEL!r}, using int8", flush=True)
+    DIFFUSION_MODEL = "int8"
+
 # (models/<subdir>, local filename, HF repo, path in repo)
 DOWNLOADS = [
-    ("diffusion_models", "minimax_h3_fl2va_pruned_nvfp4.safetensors",
-     "lilcheaty/MiniMax-H3-NVFP4", "minimax_h3_fl2va_pruned_nvfp4.safetensors"),
+    ("diffusion_models", *DIFFUSION_MODELS[DIFFUSION_MODEL]),
     ("text_encoders", "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors",
      "Comfy-Org/MiniMax-H3", "text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors"),
     ("vae", "minimax_h3_video_vae_fp16.safetensors",
@@ -616,7 +634,7 @@ def generate_video(
 
 # ── Startup ──────────────────────────────────────────────────────────────────
 
-print("[setup] ComfyUI bootstrap", flush=True)
+print(f"[setup] ComfyUI bootstrap — diffusion model {DIFFUSION_MODEL} ({UNET_NAME})", flush=True)
 ensure_comfy()
 ensure_models()
 init_comfy_nodes()