One-Click DiT Model Acceleration for ComfyUI
Quality Comparison (Z-Image-Base, 50 steps)
| w/o Cache-DiT Acceleration | w/ Cache-DiT Acceleration |
|---|---|
 |
 |
Guidance Video (Click below)
Thanks to Benji for the excellent tutorial!
🔥[2026.08] We added MiniMax-H3 support in ComfyUI-CacheDiT, validated on official T2V, I2V, and R2V workflows with same-seed comparisons, achieving 1.41-1.50x speedup while preserving native video output and stereo audio.
ComfyUI-CacheDiT brings 1.4-1.6x speedup to DiT (Diffusion Transformer) models through intelligent caching, with zero configuration required.
Inspired by llm-scaler, a high-performance GenAI solution for text, image, and video generation on Intel XPU.
| Model | Steps | Speedup | Warmup | Skip_interval |
|---|---|---|---|---|
| Z-Image | 50 | 1.3x | 10 | 5 |
| Z-Image-Turbo | 9 | 1.5x | 3 | 2 |
| Qwen-Image-2512 | 50 | 1.4-1.6x | 5 | 3 |
| Flux.2 Klein 4B | 20 | 1.67x | 4 | 2 |
| Flux.2 Klein 9B | 20 | 1.67x | 4 | 2 |
| MiniMax H3 | 20 | 1.41-1.50x | 3 | N/A (adaptive) |
| LTX-2 T2V | 20 | 2.0x | 6 | 4 |
| LTX-2 I2V | 20 | 2.0x | 6 | 4 |
| WAN2.2 14B T2V | 20 | 1.67x | 4 | 2 |
| WAN2.2 14B I2V | 20 | 1.67x | 4 | 2 |
MiniMax H3 uses residual-difference-based adaptive DBCache decisions rather
than a fixed skip_interval, so that column is not applicable to H3.
pip install -r requirements.txtClone Repository
cd ComfyUI/custom_nodes/
git clone https://github.com/Jasonzzt/ComfyUI-CacheDiT.gitFor Image Models (Z-Image, Qwen-Image Flux.2 Klein):
- Load your model
- Connect to ⚡ CacheDiT Accelerator node
- Connect to KSampler - Done!
[Load Checkpoint] → [⚡ CacheDiT Accelerator] → [KSampler]
For Video Models (MiniMax H3, LTX-2, WAN2.2 14B):
MiniMax H3:
[Load Diffusion Model] → [⚡ CacheDiT Accelerator] → [Model Sampling / Guider]
The Auto preset detects MiniMaxH3Model and enables Pattern 3 DBCache with
in-place-safe residual boundaries for its joint audio-video transformer.
For explicit H3 tuning, use the optional ⚡ CacheDiT MiniMax H3 Advanced
node in the same position. It exposes only the four H3 DBCache controls that
are useful for workflow-level quality/speed tuning; Auto remains the
recommended starting point.
The initial H3 preset is F8 / B0 / threshold 0.12 / warmup 3:
| Control | Initial value | Meaning and initial-setting rationale |
|---|---|---|
fn_blocks |
8 | Always execute the first 8 blocks. This gives the residual comparison an early, current-step signal while leaving most blocks eligible for reuse. |
bn_blocks |
0 | Do not recompute tail blocks on a cache hit. This is the initial performance-oriented setting; raise it if a workflow needs more conservative tail refinement. |
residual_diff_threshold |
0.12 | Reuse the middle-block residual when the normalized residual difference stays below 0.12. Lower values are more conservative; higher values usually increase cache hits and quality risk. |
warmup_steps |
3 | Fully compute steps 0, 1, and 2; adaptive cache reuse becomes eligible at step 3. |
These values are a validated integration baseline, not a model-wide optimum. T2V, I2V, and R2V workflows can respond differently to the same cache settings. Tune one control at a time and compare the same workflow, seed, steps, resolution, frame count, audio settings, and output codec.
LTX-2 Models:
[Load Checkpoint] → [⚡ LTX2 Cache Optimizer] → [Stage 1 KSampler]
WAN2.2 14B Models (High-Noise + Low-Noise MoE):
[High-Noise Model] → [⚡ Wan Cache Optimizer] → [KSampler]
[Low-Noise Model] → [⚡ Wan Cache Optimizer] → [KSampler]
Each expert model gets its own optimizer node with independent cache.
| Parameter | Type | Default | Description |
|---|---|---|---|
model |
MODEL | - | Input model (required) |
enable |
Boolean | True | Enable/disable acceleration |
model_type |
Combo | Auto | Auto-detect or select preset |
warmup_steps |
Integer | 0 | Generic node: 0 uses the selected preset |
skip_interval |
Integer | 0 | Generic node: 0 uses the selected preset |
print_summary |
Boolean | True | Show performance dashboard |
That's it! The generic node configures technical parameters automatically. Use the H3 Advanced node only when you need explicit H3 workflow tuning.
Caching Logic:
# After warmup phase (first 3 steps)
if (current_step - warmup) % skip_interval == 0:
# Reuse cached result
result = cache
else:
# Compute new result
result = transformer.forward(...)
cache = result.detach() # Save to cacheBased on cache-dit by Vipshop's Machine Learning Platform Team.
Built for ComfyUI - the powerful and modular Stable Diffusion GUI.
Note for LTX-2: This audio-visual transformer uses dual latent paths (video + audio). Use the dedicated ⚡ LTX2 Cache Optimizer node (not the standard CacheDiT node) for optimal temporal consistency and quality.
Note for WAN2.2 14B: This model uses a MoE (Mixture of Experts) architecture with High-Noise and Low-Noise models. Use the dedicated ⚡ Wan Cache Optimizer node (not the standard CacheDiT node) for best results.
Other DiT models should work with auto-detection, but may need manual preset selection.
A: Currently, only Z-Image-Turbo (9 steps) has been tested and verified. Other low-step distilled models require further validation.
For extremely low step counts (< 6 steps), the warmup overhead significantly reduces the benefit - sacrificing quality for minimal speed gains is generally not worthwhile in such cases.
A: Simply set enable=False in the node and run it once. This will cleanly remove the CacheDiT optimization from your model without requiring a restart.
A: This usually means:
- Model not properly detected - try manual preset selection
- Inference steps too short (< 10 steps) - warmup takes most steps
- Check logs for "Lightweight cache enabled" message
A: Properly configured (default settings), quality impact is minimal:
Star ⭐ this repo if you find it useful!
