Updated 2026-09 · Organized as symptom → cause → fix, all based on community testing
| Symptom | Cause | Fix |
|---|
| Output is solid black |
The INT8 ConvRot VAE needs a newer kernel; manual fp16 conversion overflows on some cards |
Upgrade ComfyUI to ≥0.31 (which includes int8 VAE support); cards that black-screen with --fp16-unet (V100 and others) need a dedicated fp16 fix patch |
| Gray noise / snow on long clips |
SageAttention has a silent-noise bug on very long sequences (above roughly 160k tokens) |
For long clips, switch to the comfy-kitchen attention backend, or the Patch Sage node from KJNodes, or shorten the duration and generate in segments |
| Blurry video / seam blocks |
When VRAM runs short during VAE decode, it silently falls back to tiled decoding |
Lower the peak demand (close other VRAM-hungry programs), or just drop the resolution; don't force high resolutions for rollouts |
| Symptom | Cause | Fix |
|---|
| Chopped audio / pops / wrong words |
Video and audio run on two separate sampling schedules; with too few steps the audio gets oversampled (especially at 4 steps) |
Switch to 6-8 steps, or use a sampler that supports dual-clock scheduling (the larry LoRA must be paired with a dual-clock sampler) |
| Recommended audio sampling parameters |
The official template defaults don't fully match the distilled LoRA |
A common combo: video sigma shift 12 / audio shift 3; sampler res_multistep or euler + simple |
| Audio completely garbled on older versions |
Before ComfyUI 0.30, audio was passed through the sampler differently |
Upgrade to ≥0.30 (the first fix landed there); going straight to 0.34+ is recommended |
| Symptom | Cause | Fix |
|---|
| Out of VRAM (OOM) |
Weights + activations + cache exceed the limit |
Add --lowvram; reduce resolution/duration; switch to lower quantization (GGUF Q3/Q4, W4A8); close VRAM-hungry programs such as browser hardware acceleration |
| Speed far below expectations |
Attention backend disabled / LoRA re-patched at every step / spilling to the pagefile |
Use Sage or Kitchen attention; with low VRAM + quantization + LoRA, prefer "weights with the LoRA fused in" (skips the per-step patch); check Task Manager to confirm there's no heavy pagefile I/O |
| Each clip slower than the last, ending in a crash |
ComfyUI slows down (~1.8×) before failing, typically during long uninterrupted sampling runs |
Restart ComfyUI as soon as you notice the regular slowdown; split batch renders into separate jobs and restart once in between |
| VRAM blows up after switching LoRA nodes |
Old allocations aren't released when switching between Turbo sampling nodes and native nodes |
Restart ComfyUI before switching (the most frequently reported community pitfall) |
| Symptom | Cause | Fix |
|---|
| Crash loading large models on Windows |
mmap-reading safetensors files over 20GB crashes |
Launch with --disable-mmap (ComfyUI ≥0.31) |
| LoRA loads but has no effect |
H3 LoRAs in diffusers format don't work directly in ComfyUI (different key layout) |
Use a verified converted version (qkv fused, fc1 halves swapped); don't use the original diffusers file |
| GGUF shape errors / conflicts with memory-optimization nodes |
Some optimization nodes read the GGUF container shape as the weight shape |
Don't stack these memory-optimization nodes on GGUF setups; use the platform's built-in offload strategy |
| Can't find the MiniMax H3 nodes |
Version too old |
Upgrade ComfyUI to ≥0.30 (native support since 0.30.0) and pick the H3 workflow from Templates → Video |
| VRAM peak stays pinned at the ceiling |
ComfyUI's dynamic VRAM runs up against the limit by design; this is normal behavior |
Don't panic; what matters is "does it hit the pagefile / does it OOM," not the peak number itself |