- MiniMax H3 cache settings directly control reroll speed and VRAM efficiency in ComfyUI
- RAM pressure caching skips the 32B text encoder on seed rerolls, saving massive time
- Avoid
cache noneas it discards encoded conditioning and forces full reprocessing - Easy Cache offers optional step-skipping acceleration but alters final visual output
- 8GB VRAM is viable when dynamic VRAM streaming and correct quantization are applied
Understanding MiniMax H3 Cache Architecture
MiniMax H3 is a massive AI video generation model with no lightweight variant. The lightest usable stack requires approximately 42.5 GB of model weights, including 21 GB for the diffusion model, 15.7 GB for the text encoder, and 5.8 GB for the VAEs. Running this on an 8 GB GPU relies entirely on ComfyUI's dynamic VRAM system, which streams weights layer-by-layer from your storage drive.
Because of this streaming behavior, the system is bandwidth-bound rather than compute-bound. Your storage speed and bus bandwidth determine the generation time. This is where caching becomes critical—retaining processed data between runs prevents unnecessary recomputation.
Video Highlights:
- Full workflow demonstration running MiniMax H3 on 8GB VRAM
- Explanation of dynamic VRAM streaming and bandwidth bottlenecks
- Breakdown of why RAM pressure caching is essential for speed
- Comparison of INT8 vs FP8 quantization on older GPU architectures
- Step-by-step ComfyUI node configuration and preset management
Never add the cache none flag to your launch arguments. The default RAM pressure caching is what retains your encoded conditioning between runs. Without it, ComfyUI forces the 32 billion parameter text encoder to reprocess from scratch every single time.
Cache Types and Their Functions
| Cache Mechanism | Function | Impact on Speed | Impact on Quality |
|---|---|---|---|
| RAM Pressure Cache (Default) | Stores encoded conditioning between runs | Massive speedup on rerolls | No impact |
| Easy Cache (Optional) | Skips transformer steps when drift is low | Moderate speedup per generation | Slight changes to output |
| Cache None (Flag) | Disables all caching | Severe slowdown | No impact |
Essential ComfyUI Launch Arguments
Configuring the correct launch arguments is the foundation of a stable MiniMax H3 cache workflow. Many commonly recommended flags are actively harmful when using dynamic VRAM streaming on low-memory cards.
The most important rule: do not add lowvram. This flag is explicitly ignored when dynamic VRAM is active on modern Nvidia setups with native Windows and Torch. Its own help text confirms this, meaning it wastes time and accomplishes nothing.
Set reservevram to 0.9. This leaves a small headroom so Windows background processes are not squeezed out, preventing system instability during heavy generation loads without significantly impacting model performance.
Recommended Launch Flags
| Flag | Recommended Value | Reason |
|---|---|---|
--reservevram | 0.9 | Prevents Windows system crashes |
--cache-classic or default | Enabled | Retains conditioning for rerolls |
--lowvram | Do NOT use | Ignored by dynamic VRAM, wastes time |
--cache-none | Do NOT use | Destroys reroll speed entirely |
Required Custom Nodes and Packages
| Component | Purpose | Installation |
|---|---|---|
| KJ Nodes (Sage Attention Patch) | Attention optimization for speed | ComfyUI Manager |
| Sage Attention (Python package) | Backend attention acceleration | pip install into embedded interpreter |
Quantization Choices for Cache Efficiency
Quantization format selection directly impacts how efficiently the MiniMax H3 cache system operates. The correct choice depends entirely on your GPU's Streaming Microarchitecture (SM) version, not on general advice.
A common misconception is that FP8 models always provide a speed boost. This advice only holds true for RTX 40-series or 50-series cards. On older architectures, FP8 is actively slower than INT8 because it gets emulated in software rather than running natively on tensor cores.
INT8 Quantization
- Requires SM 7.5 or higher
- Runs natively on tensor cores
- Best choice for RTX 20/30-series
- No emulation overhead
FP8 Quantization
- Requires SM 8.9 or higher
- Native on RTX 40/50-series
- Emulated on older cards
- Slower than INT8 if emulated
NVFP4 (Text Encoder)
- Used for Qwen 3 VL encoder
- Emulated on most cards
- Runs once per prompt, then cached
- Negligible cost per step
If your card lacks the required SM version for a quantization format, it does not fail—it emulates. Emulation performs the full computation in software and discards the speed benefit, making it slower than a lower format that runs natively.
Benchmark Comparison on SM 86 (RTX 30-series)
| Format | Big Matmul Time | Native Support | Verdict |
|---|---|---|---|
| BF16 Dense | 6.69 ms | Yes | Baseline |
| INT4 | 1.28 ms | Yes | Fastest |
| NVFP4 | 6.20 ms | Emulated | No speed gain over BF16 |
Step-by-Step Cache Configuration
Install Sage Attention
Install the Sage Attention Python package into your ComfyUI embedded interpreter. This is the single required custom package for the entire workflow. Without it, the Sage Attention Patch node from KJ Nodes will not function.
Configure Launch Arguments
Remove any lowvram or cache none flags from your startup script. Add --reservevram 0.9 to ensure Windows stability. Verify that default RAM pressure caching remains active so conditioning persists between seed rerolls.
Load the Model Chain
Wire the diffusion loader (INT8 file) into Sage Attention, then into Sigma Shift, then into Easy Cache. Feed the output to both the scheduler and the guider. The order matters—Sigma Shift must sit before the scheduler because the scheduler reads model sampling.
Set Up the Preset Node
Use the integer preset node to control resolution and duration simultaneously. Presets 1 through 6 offer pre-balanced combinations. Preset 0 switches to manual mode for custom width, height, and frame count inputs.
Enable Easy Cache (Optional)
Press Control B to enable Easy Cache for exploration runs. Set the reuse threshold to 0.2, start percent to 0.15, and end percent to 0.95. Monitor the console with verbose mode to see how many steps are skipped. Disable it for final renders.
Start with Preset 3 at 14 steps to find your framing and motion. Keep the prompt locked and reroll seeds—this is nearly free because the text encoder output is cached. Once satisfied, increase steps to 20 and move to Preset 4 or 6 for the final high-quality render.
Easy Cache Deep Dive
Easy Cache is an optional acceleration mechanism built into the MiniMax H3 workflow chain. It predicts how much each sampling step would change the output. When the predicted drift falls below a defined threshold, it skips the transformer entirely and reuses the previous step's result.
This system costs almost no additional memory, making it safe for low-VRAM setups. However, it does alter the final visual output, which is why it remains bypassed by default.
Easy Cache Parameters
| Parameter | Recommended Value | Function |
|---|---|---|
| Reuse Threshold | 0.2 | Speed vs quality dial; higher means more skips |
| Start Percent | 0.15 | Never skips early steps that set composition |
| End Percent | 0.95 | Never skips late steps that add fine detail |
| Verbose | True | Logs skipped steps to console for monitoring |
Turn Easy Cache on while exploring different prompts and compositions. The speed savings let you iterate faster. Once you lock in a prompt and seed you like, bypass Easy Cache (Control B) for the final render to ensure maximum visual fidelity.
Prompting and Generation Limits
MiniMax H3 uses a joint audio-video generation pipeline. A single latent is decoded through two separate VAEs—one for video frames and one for stereo audio. This architecture fundamentally changes how you must write prompts.
Write one continuous block of text covering visual look, camera shots, and audio design. H3 handles shot lists well—you can literally write "shot one, shot two, cut to" and the model follows the structure.
The workflow uses a Basic Guider instead of a CFG Guider. This means there is no CFG scale and no negative prompt field. Anything you do not want in the video must be stated as a positive instruction: "no cuts, no dissolves, no text overlays."
Duration and Resolution Presets
| Preset | Resolution | Duration | Use Case |
|---|---|---|---|
| 1 | 864x480 | 15 seconds | Maximum length, lower quality |
| 3 | 1344x768 | 5 seconds | Balanced exploration |
| 4/6 | Higher | 5 seconds | Final high-quality render |
| 0 | Manual | Manual | Custom configurations |
Maximum Generation Limits
| Constraint | Value | Reason |
|---|---|---|
| Max frame count | 362 frames (~15.08s) | Model's trained maximum |
| 20-second target | Not possible | 481 frames exceeds memory ceiling |
| Token limit at 480p | ~57,000 tokens for 20s | Past memory ceiling |
MiniMax H3 Cache Setup Checklist:
- Install Sage Attention Python package in embedded interpreter
- Remove lowvram and cache none from launch arguments
- Set reservevram to 0.9 for Windows stability
- Verify INT8 quantization matches GPU SM version
- Confirm KJ Nodes Sage Attention Patch is wired in model chain
- Write audio descriptions in every prompt block
Frequently Asked Questions
Q: What does the MiniMax H3 cache system actually do?
The default RAM pressure caching stores your encoded conditioning between generation runs. This means when you reroll a seed with the same prompt, ComfyUI skips reprocessing the 32 billion parameter Qwen 3 VL text encoder entirely, providing the single biggest speedup available in the workflow.
Q: Should I use FP8 quantization for better cache performance?
Only if you have an RTX 40-series or 50-series card with SM 8.9 or higher. On older cards like RTX 20 or 30-series, FP8 is emulated in software and is actually slower than INT8, which runs natively on tensor cores. Check your GPU's SM version before choosing a quantization format.
Q: Why can't I generate clips longer than 15 seconds?
The model's trained maximum is 362 frames, which equals approximately 15.08 seconds at 24 FPS. Beyond that range is untested. Additionally, 20 seconds would require around 481 frames and approximately 57,000 tokens even at 480p, which exceeds the memory ceiling regardless of cache configuration.
Q: Does Easy Cache improve quality or just speed?
Easy Cache only improves speed. It works by skipping transformer steps when the predicted output drift is below a threshold. This means it changes the final visual output slightly. Use it during exploration and iteration, but disable it for final renders where maximum fidelity matters.