Flux Tools (Depth/Canny/Redux/Fill) Model Architecture & Placement

Black Forest Labs' four Flux Tools models are not a uniform family — each is a distinct model type with its own installation folder and conditioning mechanism:

Flux.1 Depth (LoRA) — depth-guided generation. Despite behaving like a ControlNet, it's implemented as a LoRA applied to the base Flux model. Install into models/loras; requires a Depth Anything preprocessor for the conditioning image. LoRA weight tuned above 1.0 (~7.5), Flux guidance ~10.
Flux.1 Canny (LoRA) — edge-guided generation, also a LoRA (not a ControlNet), reusing the Depth workflow structure with a Canny Edge preprocessor instead. LoRA weight tuned below 1.0 (~0.75), Flux guidance ~10.
Flux.1 Redux (dev) — a 129MB style model (not a LoRA), installed into models/style_models. Produces prompt-free stylistic variations of an input image; requires a separate CLIP Vision encoder (sigclip/patch14, in models/clip_vision) loaded via a Load Style Model node.
Flux.1 Fill (dev) — a full ~24GB Flux checkpoint with inpainting baked in, installed into diffusion_models. Driven by ComfyUI's built-in Mask Editor ('Open in Mask Editor' → paint → 'Save to Node'); Flux guidance ~30.

Key surprise: Depth/Canny are LoRAs, not ControlNets, which changes both where the file goes (models/loras) and which knob tunes strength (a LoRA-weight slider, not a ControlNet-strength slider). Guidance values don't carry a single default across the family — Fill needs roughly 3x the guidance of Depth/Canny, and LoRA weight is tuned in opposite directions for Depth (above 1.0) vs. Canny (below 1.0).

Source: "FLUX TOOLS - Run Local" tutorial (filed 2026-08-06).

Из тем: Flux & ComfyUI: Local Model Tuning