KreaPhoton:
There is a familiar ritual in the world of AI image generation. A new model drops, and within hours the community ports over its favorite bag of tricks ā the schedules, the CFG tweaks, the ancestral noise settings that worked on the last dozen checkpoints. Most of the time, it mostly works. And most of the time, "mostly works" is where everyone stops.
KreaPhoton starts where that stops.
Built specifically for Krea 2 Turbo ā a CFG-distilled, rectified-flow diffusion transformer on the 16-channel Wan21 latent ā the pack begins from an uncomfortable premise: on this model, most of the inherited folklore either does nothing or actively breaks the image. Naive classifier-free guidance, SD-tuned schedules, default ancestral sampling ā a distilled turbo model shrugs them off or falls apart. So instead of porting tricks, the authors measured. Over four hundred real generations, a cycle of experiments and hand-derived mathematical models mapped what this specific model actually responds to.
guidance scheme applies real CFG only inside the narrow band where it earns its keep ā sparing you the doubled compute of full-trajectory guidance. Every default value in the pack carries a label: measured fact, protocol-validated, or ā refreshingly ā untested.
That last word is the tell. KreaPhoton's most distinctive feature isn't a schedule or a guidance mode; it's an allergy to comfortable claims. When the "variety" control underdelivered, the team didn't quietly ship it and move on. They ran a pre-registered validation protocol, built a composition-versus-texture metric because the standard one was lying to them, and proved something genuinely useful and slightly deflating: on this architecture, a variance-preserving latent nudge can move texture but not composition ā the two are welded together through the initial noise field. Rather than bury the finding, they turned it into a feature. The new composition blend spherically interpolates between two seeds' noise, walking a coherent, photoreal path through composition space ā the honest lever the old knob could never be.
Around the math sits the ergonomics you actually want: live per-step previews that work regardless of your server settings, an on-node thumbnail in the style of the community's beloved efficiency nodes, and a minimal "seed / preset / variety" interface that hides the calibrated complexity underneath.
None of this is hype-shaped. There is no promise that KreaPhoton will make every image a masterpiece. What it offers instead is rarer in this space: a set of tools where you can trace every knob back to a number, a generation, or an honest admission of ignorance. In a field that runs on vibes and screenshots, that is close to a radical position.
KreaPhoton is open source under the MIT license and available now for ComfyUI.
ā github.com/Kostik2702/ComfyUI-KreaPhoton
Examples:
