add ComfyUI section: base checkpoint requirement (int8_convrot incompatible), setup, strengths

README.md

@@ -100,6 +100,43 @@ image-to-video LoRA siblings:
 480P / 4:3 is the period-correct sweet spot; the style carries fine to higher
 resolutions if you want clean pixels of a dirty tape.
 
+## ComfyUI
+
+`vh5tape.safetensors` loads in ComfyUI as-is — no conversion needed. It uses the
+ComfyUI-native key layout (`diffusion_model.blocks.N.attn.{qkv_proj,out_proj}`,
+fused QKV, alpha = rank), so all 104 modules load with no warnings.
+
+**Base checkpoint — this is the thing that matters.** Use a non-pruned,
+non-rotated base:
+
+- ✅ `minimax_h3_fl2va_bf16.safetensors` (recommended)
+- ✅ `minimax_h3_fl2va_pruned_fp8_scaled.safetensors`
+- ❌ `*_int8_convrot`, `nvfp4`, `w4a8` — these store weights in a rotated /
+  quantized basis. The LoRA loads without any error and then produces warped
+  faces, melting limbs and disappearing objects. **The default checkpoint in the
+  official ComfyUI H3 tutorial is `minimax_h3_fl2va_pruned_int8_convrot` — you
+  must change it.** (Not specific to this LoRA — MiniMax-H3 LoRAs in general are
+  made for fp16/fp8 bases.)
+
+**Setup**
+
+1. Put `vh5tape.safetensors` in `ComfyUI/models/loras/`.
+2. `UNETLoader` (bf16 or fp8_scaled fl2va) → `LoraLoaderModelOnly` →
+   `MiniMaxH3ImageToVideo` / your sampler.
+3. Strength **1.0** (useful range 0.7–1.2).
+
+**Stacking with a turbo LoRA:** drop this one to **0.4–0.6** — the two
+perturbations add, and 4-step turbo has little headroom. (The `turbo_mode`
+checkbox occupies its own LoRA slot; add a second `LoraLoaderModelOnly` for
+this one.) 4-step turbo also softens the tape grain — ~20–25 full-precision
+steps show the style best. Stay near 1 MP (very low resolutions are unstable)
+and keep prompts to a single shot.
+
+`vh5tape-comfyui.safetensors` is an optional convenience build: bit-identical
+weights plus explicit `.alpha` tensors (= rank, scale 1.0) and the base-model
+requirement in its metadata — same look, just harder to mis-load in
+third-party loaders.
+
 ## Training
 
 Trained with fal's MiniMax H3 t2v trainer (rank 32, 5,000 steps, 4:3, joint
@@ -114,6 +151,7 @@ A/B testing; more steps past 5k made the damage *tamer*, not heavier.
 | file | what |
 |---|---|
 | [`vh5tape.safetensors`](./vh5tape.safetensors) | the LoRA, rank 32, 5,000 steps |
+| [`vh5tape-comfyui.safetensors`](./vh5tape-comfyui.safetensors) | same weights bit-exact + explicit alpha tensors, for ComfyUI (see the ComfyUI section) |
 | [`PROMPTS.md`](./PROMPTS.md) | the exact prompts + seeds behind every sample above |
 
 Made for **H3-TV** — a 1970s portable CRT being rebuilt to play an endless,