Fizgig H3 Tweaks

MiniMax H3 renders in Turbo mode run without CFG. That's the point of a distilled model - the guidance is baked into the weights, so you feed it a prompt and it does what it was trained to do. Which also means there's no dial. When skin comes out slightly plasticky, or the things you named land softly, you have nothing to turn.

Fizgig H3 Tweaks is one small model-patch node that gives you four. It changes what H3's blocks write while they sample - no training, no second pass, no extra model files.

What it actually does

This is not a detailer. A detailer crops a region, upscales it, runs the model again and pastes it back. This never leaves the sampler: the detail is shaped as the latent forms, not bolted on afterwards.

From the source: it registers a replace-patch on H3's transformer blocks. Each patched block runs normally, but the node grabs the update it wanted to write to the video tokens, splits that update into frequency bands with a Gaussian blur, and adds a scaled copy of one band back. It's unsharp masking on the model's internal update rather than on pixels - d − blur(d) is fine texture, blur(d,1) − blur(d,3) is local contrast, blur(d,3) is the broad layout. It reads your sampler's step count every step, so "the last half" means whatever you're running.

Detail sits on blocks 40–49 over the last half of the steps; Scene Variation on blocks 20–49 over the first third. Text and audio rows are never touched, and everything at 0 leaves the model exactly as it was.

The inputs that matter

model - the H3 model with your LoRAs already applied. Order:

Load Diffusion Model → LoraLoaderModelOnly → Fizgig H3 Tweaks → BasicGuider / BasicScheduler → SamplerCustomAdvanced

Detail & Contrast (high_freq_detail, default 0.15, −1 to 1). Positive is crisper and punchier: more pores, freckles, lashes, deeper shadows. 0.15 is the clean default; past that it overbakes. Negative softens - smoother skin, lifted shadows.

Detail & Contrast mode (detail_mode) only matters when the slider isn't zero. Leave it on stable across frames for clips: it keeps only the detail that's identical in every frame, measured at +7% shimmer versus +20% for per frame. Per frame is fine for stills, shimmery on video.

Scene Variation (composition, default 0, ±0.5). A sub-seed - it nudges the layout on the opening steps while the overall look stays put. For the render that's 95% there. Same settings, same result, so it's safe to sweep: try ±0.1 to ±0.3.

Prompt Strength (prompt_strength, default 0, −0.5 to 3). This one works in attention, not in the update: the text tokens' values are scaled by 1 + your value, so every video and audio token draws proportionally more from the prompt. Start around 0.2 and step up.

report prints a console line per step and block. Switch it on once to check your step windows land where you expect, then turn it off. One output, model, back into the sampler chain.

Install

No Python dependencies (pyproject.toml declares an empty list) and no model downloads. Search Fizgig H3 Tweaks in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/shootthesound/ComfyUI-Fizgig-H3-Tweaks

Restart ComfyUI; it's under the Fizgig category. It rides ComfyUI's built-in MiniMax H3 support (it imports comfy.ldm.minimax.model), so you need a ComfyUI new enough to have that. There's an example workflow in the repo - ComfyUI's own MiniMax H3: Text to Video template, flattened, with the node between the Turbo LoRA and the sampler at 8 steps.

Where people get burned

• Wrong model. Anything that isn't H3 raises only knows MiniMax H3 - connect an H3 model. It only recognises the H3 class.
• Block-skip conflicts. Nodes that replace H3's blocks - block-skipping, token routing - occupy the same slots, and ComfyUI allows one replace-patch per block. The tweaks take over the blocks they share with those nodes, and print a note when they do. Use one or the other there.
• Long renders with no Turbo LoRA. Detail covers the last half of whatever step count you run, so at 20 steps it's acting on ten. Use a lower value; the node was tuned on 6–8 step Turbo.
• Two nodes chain additively. Supported, and rarely what you want.
• Overcooked detail. The author tried a whole-update gain and dropped it - grainy by 0.2. Detail & Contrast is the survivor, but +0.6 is the far end of "punchy," not a target.

Know before you start: H3's own weights are territory-restricted (the Community License excludes the US, EU, UK and South Korea). That's a MiniMax thing, not this node's.

Experimental, and it says so on the tin - the default is a mild 0.15, not something dramatic. Start there, change one dial at a time, keep the seed fixed so you can tell if a change helped.