MiniMax H3 Image • Detail Tone Lock

The detail refiner in this pack is a real second diffusion pass - it sharpens and rebuilds fine detail - but every generative edit pass drifts something. With Qwen Image Edit 2511 the typical casualty is broad lighting and color: the eyes and fabric get crisper, and somehow the whole scene gets warmer and slightly off. H3DetailToneLock is the fix. It treats the original H3 image as the authority on tone and only lets the refined image contribute the high-frequency detail you actually wanted.

What it actually does

The name sounds like marketing, but it's honest signal processing. The node does frequency separation, not masking and not an upscaler. It blurs both images with a Gaussian at detail_radius to isolate their low-frequency "tone" layers, then it restores the source's low-frequency lighting and color into the refined image, and finally mixes that tone-locked result back into the source. The math is straightforward (source plus a fraction of the difference), but the effect is that you keep Qwen's new micro-detail while H3 stays authoritative for dimensions, broad lighting, and color.

Where this fits: in the optional H3_DETAIL_REFINER workflow. Qwen runs on a two-megapixel working copy - that controls model cost only - and the original H3 image goes directly into source_image. That's why the saved result always has your original width and height, no upscale required. If the refined image doesn't match the source canvas, the node bicubically resizes it; if one refined frame is fed against a multi-frame batch, it expands it. You don't need to babysit dimensions.

The inputs that matter

The three inputs that actually matter:

• source_image - your original H3 still. It owns the composition, lighting, and color.
• refinement_strength - how much of the refined result gets mixed in. Default 0.45 in the node (the bundled workflow ships 0.55). Lower it if you see identity or geometry drift; it's the closest thing to a "how bold is this edit" dial.
• detail_radius - the pixel radius separating "broad tone" from "fine detail." The 16 default suits roughly one-megapixel images. Push it up if the refiner is still washing out large-area color.

tone_lock (default 0.85) is the restoration strength: 0 turns restoration off, 1 locks broad tone completely. The single image output is the refined image at the exact source size.

Install and the refiner stack

Install is the pack plus the Qwen refiner stack: the qwen_image_edit_2511_int8_convrot.safetensors diffusion model, the Qwen 2.5 VL text encoder, qwen_image_vae.safetensors, and the four-step Lightning adapter into models/loras/ - all covered by the official ComfyUI Qwen Image Edit guide. The pack itself has no extra Python dependencies and needs ComfyUI 0.30.0 or newer; install it via ComfyUI Manager (search MiniMax H3 Image Studio) or git clone https://github.com/astropuzzo/ComfyUI-MiniMax-H3-Image-Studio.git into custom_nodes, then restart.

The one trap to avoid

The one trap the README calls out: don't append your full H3 generation prompt to the refiner. Scene descriptions encourage reconstruction - a generic, preservation-first prompt is intentional. Describe a specific defect only when it needs repair. And a note on expectations: Detail Tone Lock restores tone, but it can't un-break an identity that the refiner already changed; that's what lowering refinement_strength is for.