The root templates ran on SDXL base, where every gallery style washed out. A
full compiled style is ~480 tokens and SDXL's CLIP attends to ~77 per chunk, so
the style was truncated before the sampler saw it. No checkpoint swap fixes
that, Juggernaut included: it is an SDXL finetune with the same encoder.
Root set now runs FLUX.1 dev (Comfy-Org fp8 all-in-one, since BFL's own file is
UNET-only and CheckpointLoaderSimple cannot load it), compiled from the flux
target, whose T5 encoder takes the whole prompt. FluxGuidance 3.5, CFG 1.0, and
ConditioningZeroOut on the negative input: FLUX is guidance distilled and has no
negative channel, exactly like Z-Image. Uses EmptySD3LatentImage because FLUX
works in a 16-channel latent space.
The two sets load different gallery styles on purpose. How well a style demos is
a property of the style and the model together: Z-Image reproduces painterly
technique and keeps the Renaissance style, FLUX is strongest at light and
tonality and takes a noir one.
Also:
- Adds the reference-images template the root set was missing, and renumbers
Facets to 05, so both sets are 01-05 with matching names. Both sets are now
pinned by test against a shared story list.
- Model download metadata moves to each node's properties.models, where ComfyUI
reads it, so a missing checkpoint offers a download instead of being silently
substituted.
- The Load node names the style its slug resolves to, on a read-only row under
style_ref, via a new /styleref/resolve route.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
