ComfyUI Sketch Renders: Nodes and Models

Which nodes and models fit a sketch?

A sketch enters the process as an image. A preprocessor extracts a hint; the compatible model uses that hint during generation. The ControlNet paper documents edge, depth and other spatial controls. It does not establish one sampler recipe for every architectural drawing. Treat the choices below as a file reference, with public workflow examples linked in the FLUX section.

The auxiliary preprocessor README maps these line nodes to SD1.5 ControlNets. Its nodes produce hint images; they do not supply the control weights.

Need to adjust Canny thresholds? Use its own node. The pack's all-in-one preprocessor does not expose those controls. Inspect the resulting hint before moving on; that is the image you are asking the model to follow.

Keep the model family explicit. Comfy's FLUX ControlNet guide requires a control model compatible with the base model. An SD1.5 Scribble weight does not become a FLUX model because its input is a drawing. The FLUX and SDXL model-routing guide covers the broader family choice.

FLUX has its own structural files

BFL's structural-conditioning documentation lists Canny for edge maps and Depth for depth maps, each as a full model and a LoRA for FLUX.1 [dev]. That set contains no Scribble model. Do not build a shopping list around an official FLUX Scribble file that this release does not provide.

The ComfyUI FLUX examples give these filenames and folders, plus importable example workflows:

The LoRA versions apply to FLUX.1 [dev]. The regular setup also needs text encoders and a VAE; follow the linked example's file list. Its single-file fp8 checkpoint is a separate route: it uses Load Checkpoint, with CFG set to 1.0 in that example. That instruction is specific to the checkpoint, not an architectural rendering preset.

Start from the public example that matches your files. Save the hint image alongside the generation so you can compare the intended lines with the output. The stage-checkpoint guide offers a way to organise those comparisons. If you plan another generation pass, the second-pass control guide addresses that separate decision. Neither step needs an invented universal strength or denoise value.

What does the workflow JSON contain?

Comfy's workflow JSON schema 0.4 describes a graph using nodes and links. At the root it requires last_node_id, last_link_id, nodes, links and version.

Each node requires id, type, pos, size, flags, order, mode and properties. Its inputs, outputs and widgets_values fields are optional in this schema. Widget values may be an array or an object.

The optional models array records file information: each entry requires name, url and directory. That is model metadata, not the weights themselves. When sharing a graph, include the exact model names and the node packages it depends on. A JSON file with an ambiguous checkpoint name leaves the next reader guessing.

Check the licence before the download

The FLUX.1 [dev] model card places its weights under a non-commercial licence and treats output use separately under those terms. The Canny and Depth releases are also labelled non-commercial. BFL offers a separate commercial weight licence. Downloadable weights are not blanket permission for commercial use.

An upscaler has its own terms. If a shared graph names 4x-UltraSharp, its author identifies it as a 4x ESRGAN model under CC-BY-NC-SA-4.0. Check that restriction before treating it as a default for paid work. This is a licence check, not a recommendation based on a render test.

An image cannot certify the drawing

FLUX.1 [dev] generates images; BFL's structural variants guide those images with edges or depth. That makes a sketch render a visual proposal, not a measured drawing or a BIM model. Keep dimensions in the design document. A convincing facade is still an image.