Deploy Flux Kontext

Cause: fp8_e4m3fn_scaled checkpoints carry per-block scale factors that preserve dynamic range, matching bf16 quality for many modern diffusion/video models at ~half VRAM. Plain (unscaled) fp8_e4m3fn CAN visibly degrade — that's why the scaled repacks exist.

Fix: Prefer the 'scaled' fp8 variant; don't reflexively pick bf16. Do NOT generalize the claim to unscaled fp8.

Cause: FLUX.2 dropped FLUX.1's dual T5-XXL+CLIP_L for a single Mistral-3-Small encoder (18GB fp8 / 36GB bf16) — and the smaller Klein 4B/9B variants use a Qwen-3 encoder instead. The encoder stack isn't stable across versions or even size tiers of one release.

Fix: Match encoder files to the exact variant (mistral*3_small_flux2 for FLUX.2 dev; qwen_3*\* for Klein). Don't reuse the previous version's encoders.

Cause: Z-Image Turbo is a photoreal-leaning generalist, not an anime model; Z-Anime (SeeSee21 fine-tune) is a separate community model on the identical architecture and loads via the same zimage deploy path. Encoder architecture is Qwen-3-4B, VAE = Flux's ae.safetensors (335MB). fp8 diffusion 6.15GB (fits 12GB), bf16 12.3GB; GGUF Q3-Q8 for low VRAM.

Fix: A Turbo base and its fine-tunes share architecture + VAE (and the Qwen-3-4B encoder architecture — a fine-tune may ship its own tuned encoder weights, e.g. Z-Anime's qwen_3_4b-engineer-v4-fp8), so one zimage deploy path serves both.

Cause: ComfyUI-GGUF's UnetLoaderGGUF reads a dedicated 'unet_gguf' folder key (fed from both unet and diffusion_models). models/unet is the reliably-covered location; the equivalent .safetensors goes in models/diffusion_models.

Fix: Route .gguf diffusion weights to models/unet (canonical, always covered). A GGUF in diffusion_models may not appear depending on ComfyUI/GGUF version — don't rely on it.

Cause: A relight LoRA above strength 0 imposes a relighting color shift matching the driving scene instead of preserving source color (shipped default is 0.0; Kijai's enabled default is 1.0).

Fix: Zero the relight/style LoRA when you want source color preserved. A tinted/off-color output means suspect an auxiliary LoRA near strength 1.0.

Cause: Unpinned by default: models pull from floating HF /resolve/main (branch HEAD), custom nodes are git-clone-latest (no SHA), Docker tags are reused, and only ~95% file-size is validated, not content hash. An upstream re-upload or newer node code silently changed bytes/behavior behind unchanged names.

Fix: Pin weights to a commit/revision, nodes to a SHA, images to a digest. Verify a pinned HF URL with `curl -sI -r 0-0 -A 'Mozilla/5.0' -L '<url>'` — MUST include -L (HF /resolve/ 302-redirects to a CDN; without -L you see 302 not 206).