09:21
Flux Tools
The video demonstrates how to run Black Forest Labs' four new Flux Tools models — Fill, Redux, Depth, and Canny — locally in ComfyUI, showing that Depth and Canny are implemented as LoRAs (not ControlNets) applied to the base Flux model, Redux is a separate style model that generates prompt-free variations, and Fill is a full inpainting checkpoint driven by ComfyUI's built-in mask editor, each requiring specific model placement and parameter tuning to work correctly.
Flux released four new tools — Fill (inpainting), Redux (variation), Canny, and Depth — and the video shows how to run all of them locally in ComfyUI.
Depth and Canny are explicitly not ControlNets; they are LoRAs applied to the base Flux model, despite behaving similarly to one.
The Depth and Canny LoRAs must be downloaded after accepting a site agreement, then placed in the models/loras folder; official example workflows are loaded by downloading each page's example image and dragging it into ComfyUI.
Running the Flux Tools example workflows requires updating ComfyUI first via Manager → Update All, then restarting the app.
The Depth workflow uses the Flux.1-Depth-fp8 model, LoRA weight 7.5, Flux guidance 10, and a Depth Anything-preprocessed image as the conditioning input.
The Canny workflow reuses the same structure but with the Canny LoRA at weight 0.75 and a Canny Edge preprocessor instead of Depth Anything.
The presenter changes the default KSampler settings from 20 steps/euler/normal to 35 steps/euler/simple for better results, then upscales the low-resolution base output with Ultimate SD Upscale plus the NMKD Superscale model.
Redux (129MB) is not a LoRA — it is a style model placed in models/style_models, requires a separate CLIP Vision ('sigclip') model in models/clip_vision, uses no text prompt, and produces re-runnable variations of an input image.
Flux Fill needs the full ~24GB Flux model with inpainting baked in, placed in the diffusion_models folder; masking is done via ComfyUI's built-in Mask Editor ('Open in Mask Editor' → paint → 'Save to Node'), and it runs with Flux guidance set to 30.
Portrait results upscaled with Flux can look 'plasticky' on the face even when the overall composition is preserved.
Flux.1 Depth (LoRA) — A depth-guided image generation model that the video says is implemented as a LoRA applied to the base Flux model rather than as a ControlNet, even though it behaves like one. Apply: Download the Depth LoRA (after accepting the site's usage agreement) into the ComfyUI models/loras folder, load it with the Flux.1-Depth-fp8 base model and Dual CLIP Loader, feed it a Depth Anything-preprocessed image as the pixel input, set LoRA weight to 7.5 and Flux guidance to 10, then sample.
Flux.1 Canny (LoRA) — An edge-guided image generation model, also released as a LoRA rather than a ControlNet, that reproduces an input image's composition from a Canny edge map. Apply: Download the Canny LoRA into the loras folder, reuse the Depth workflow but swap the preprocessor for a Canny Edge node, set the LoRA weight to 0.75 (reported as better than 1.0), then sample and upscale.
Flux.1 Redux (dev) — A 129MB style/variation model (not a LoRA) that takes an input image and produces stylistic variations of it, with no text prompt involved. Apply: Place the Redux file in models/style_models (not the loras folder), load it via a Load Style Model node together with the Flux.1-Depth-fp8 base model, a CLIP Vision (sigclip) encoder, and the Dual CLIP Loader, feed in a source image, then queue repeatedly to get different variations of the same image.
Flux.1 Fill (dev) — A ~24GB full Flux model with an inpainting ('fill') capability baked in, used to replace a masked region of an image based on a text prompt. Apply: Put the full Fill checkpoint into the diffusion_models folder, load an image, right-click it and choose 'Open in Mask Editor' to paint the area to replace (adjusting thickness/opacity), click 'Save to Node', set Flux guidance to 30, and run the queue.
ComfyUI — The node-based UI the video uses to run all four Flux Tools models locally via example workflow graphs. Apply: Get the official example workflows from the comfyanonymous ComfyUI_examples page and drag each page's example image directly into the ComfyUI canvas to auto-load the matching workflow.
ComfyUI Manager 'Update All' — A built-in updater in ComfyUI's Manager panel needed to bring ComfyUI to the version required for the new Flux Tools example workflows. Apply: Open the Manager on the right side of ComfyUI, click 'Update All', watch the CMD window until it finishes, then restart ComfyUI before loading the Flux Tools workflows.
Dual CLIP Loader (clip L + T5 XXL fp16, type flux) — The text-encoder loading node used across all the Flux Tools workflows, combining a CLIP-L model and a T5-XXL fp16 model set to the 'flux' type. Apply: Load clip L and T5 XXL fp16 together in the Dual CLIP Loader node and route its output into the model/CLIP chain feeding the KSampler and VAE in each Flux Tools workflow.
Flux Guidance parameter — A per-workflow guidance-strength setting the video says needs unusually specific, model-dependent values to get correct results. Apply: Set Flux guidance to 10 for the Depth and Canny LoRA workflows and to 30 for the Fill inpainting workflow, per the video's tested values.
KSampler settings (steps/sampler/scheduler) — The video's custom deviation from the official example workflow's default sampling settings for the Depth/Canny workflow. Apply: Replace the default 20 steps / euler / normal scheduler with 35 steps / euler / simple scheduler for better results, per the presenter, while still experimenting with other values.
Depth Anything (preprocessor) — The preprocessing node used to convert a loaded image into a depth map before it is fed into the Depth LoRA workflow. Apply: Insert a Depth Anything node between the loaded source image and the pixel input of the Flux Depth LoRA setup to generate the depth-conditioning image.
Canny Edge (preprocessor) — The preprocessing node (one of several possible Canny nodes) used to convert a loaded image into an edge map for the Canny LoRA workflow. Apply: Route the source image through a Canny Edge node (or any equivalent Canny preprocessor) and connect its output image to the pixel input feeding the KSampler in the Canny LoRA workflow.
Ultimate SD Upscale (with NMKD Superscale) — An upscaling technique applied after the base Depth/Canny generation to increase resolution and detail, since the base output is described as low-resolution and low-detail. Apply: Feed the KSampler's low-res output into an Ultimate SD Upscale node configured with the NMKD Superscale model to produce a higher-resolution, more detailed final image.
CLIP Vision Loader (sigclip model) — A vision-encoder loading node required specifically for the Redux workflow, using a Google-provided model the video identifies as 'sigclip' (patch14) despite the official page calling it 'SigLIP'. Apply: Download the sigclip patch14 clip vision model into models/clip_vision and load it via the CLIP Vision Loader node feeding the Redux style-model pipeline; click 'refresh' in the loader if the model doesn't appear.
Load Style Model node (Redux) — The node that loads the Redux weights as a 'style model' rather than a LoRA, distinguishing it architecturally from the Depth/Canny approach. Apply: Place the downloaded Redux file in models/style_models and load it in the Load Style Model node, connecting it alongside the base Flux Depth fp8 model, CLIP Vision output, and Dual CLIP Loader with no text prompt needed.
Mask Editor / 'Save to Node' (inpainting mask) — ComfyUI's built-in mask-painting tool used to define the region of an image the Fill model should regenerate. Apply: Right-click the loaded image, choose 'Open in Mask Editor', paint over the area to replace using the thickness/opacity controls, then click 'Save to Node' before queuing the Fill workflow.
The video flags an architectural surprise: BFL packaged Depth/Canny conditioning as LoRAs instead of ControlNets, which changes where the files are installed and how strength is tuned (a LoRA-weight slider rather than a ControlNet-strength slider).
The guidance values needed are counterintuitive across models — 10 for Depth/Canny versus a much higher 30 for Fill — which the presenter calls 'strange' and 'pretty crazy' even while confirming they work, implying no single guidance default carries across the Flux Tools family.
LoRA weight needs opposite tuning per model: Depth is pushed above 1.0 (7.5) while Canny is pulled below 1.0 (0.75); the video presents both as empirically better than the workflow's default without explaining why.
The presenter's own KSampler tweak (35 steps/simple scheduler vs. the official 20/normal) is framed as a discovered improvement over Black Forest Labs' own example workflow defaults, not just a settings walkthrough.
Redux's CLIP Vision requirement is inconsistently named across sources — the official page calls it 'SigLIP' but the presenter says the actually-available file needed is named 'sigclip' — a naming mismatch worth flagging when sourcing the model.
«not a control net this is a Laura that is applied to the image»
— 00:34
«very important detail here you need to have the updated version of com fui»
— 00:33
«this is at 10 which is the correct value in this case it's strange but it does work»
— 00:42
«create for you A variation as you can see it works pretty well it has all of the details in there but it's a different variation of the same image»
— 01:01
«I have my fill model guided 30 pretty crazy but it works»
— 00:56
«some pretty fantastic models most of them small all size run pretty smoothly easy to understand workflow»
— 00:55
Reception
Engaged but frustrated audience struggling with technical issues, model behavior inconsistencies, and ComfyUI's complexity, with minimal enthusiastic feedback.
A concise, hands-on tutorial that clearly documents where each new Flux Tools model and support file goes and how the workflow differs from a ControlNet setup, though it presents several non-default parameter choices as working without explaining the underlying reason.
09:21