Controlled generation (ControlNet)

Generate a new image whose composition matches a source image while its style, subject, or finish changes freely. You extract a structural map (pose, Canny edges, depth, or segmentation) from the source, then condition generation on that map. This is the "same composition, new look" lever: interior redesign, sketch-to-render, and repeatable product or game-asset variations.

Inputs to collect

• The source image to take structure from (a photo, sketch, render, or existing asset). Required.
• Which structure to preserve: silhouette/outlines (Canny), human pose (pose), 3D layout/perspective (depth), or scene regions (segmentation). If unstated, infer from the goal: outlines for objects/assets, pose for people, depth for rooms/scenes.
• What changes: the new style, subject, material, or palette described as a prompt.
• How tightly to hold the structure (loose restyle vs strict trace) and how many outputs.

Models

Two pieces, both confirmed live via runware-models before calling, never hardcoded:

1. A ControlNet preprocessor (form:controlnet consumers depend on it) to build the control map:
   • Canny / edges: runware:controlnet-preprocess@canny (op:edge-detection)
   • Depth: runware:controlnet-preprocess@depth (op:depth-estimation)
   • Pose: runware:controlnet-preprocess@openpose (op:pose-estimation)
   • Segmentation: runware:controlnet-preprocess@seg (op:segmentation)
   • Soft edges: runware:controlnet-preprocess@softedge (op:edge-detection)
2. A base model + a matching ControlNet (form:controlnet) for inference. The ControlNet must match both the base family and the map type:
   • FLUX.1 [dev] base runware:101@1 with FLUX Union Pro ControlNets: Canny runware:25@1, Depth runware:27@1, Pose runware:29@1.
   • SDXL alternatives: Canny runware:20@1, Depth runware:3@1.

Confirm each AIR is live and the ControlNet is built for the base you chose. A LoRA can ride alongside to lock a style (see the game-assets-canny guide). Offer SDXL as the faster/cheaper tier and FLUX dev as the quality tier.

Workflow

Specializes the runware-run contract for a two-call control pipeline.

1. Resolve schemas. Inspect the preprocessor and the inference-model schemas via runware-run. Use only schema-valid fields.
2. Extract the control map. Run taskType: "controlNetPreprocess" with inputImage (URL or base64) and preProcessorType set to the map you want (canny, depth, openpose, seg). Keep the returned guide image.
3. Generate conditioned on the map. Run taskType: "imageInference" synchronously with your model (base), the prompt, and a controlNet array whose entry carries the matching ControlNet model, the guideImage from step 2, and weight / startStep / endStep.
4. Read the result. Images return synchronously as image URLs. Inspect for structure adherence vs creative freedom and adjust.
5. For a set, reuse the same guide image across calls and vary only the prompt so every output shares one composition.

Load references/examples.md for three worked, schema-verified recipes (Canny edge-locked restyle, OpenPose pose-driven generation, depth-guided interior restyle) with both calls and the result shape.

Technique

Grounded in the FLUX dev game-assets-canny guide.

• Extract, then condition. The map carries the structure so the prompt no longer has to. Describe the new look in the prompt, not the geometry the guide already encodes.

• Tune Canny thresholds to set how much structure survives. lowThresholdCanny catches subtle edges (more detail, more constraint), highThresholdCanny keeps only strong edges (cleaner, looser). Defaults work for most inputs. Push low for soft-transition art, push high to drop noise.

• Control the structure-vs-creativity balance with steps. startStep: 1, endStep: 10 (on a ~30-step run) is the balanced default: guidance shapes the early layout, then the model is free to invent detail. A higher startStep lets guidance kick in too late and structure adherence weakens. A higher endStep clamps the model to the source so hard that the new subject can barely emerge (good for pure restyle, bad for transformation).

• Control mode trades guide against prompt. Use balanced for a mix, controlnet to prioritize structure over the prompt, prompt to let text lead with the guide as a loose reference.

• Match map to intent. Canny for silhouettes and assets, pose for redirecting people, depth for rooms and perspective (interior redesign, sketch-to-render), segmentation to keep scene regions while restyling each.

  Pick the pipeline by filling this routing template, then send the two calls:

  control type: <canny | openpose | depth | seg>
  source image: <the URL or base64 to take structure from>
  preprocessor:  <runware:controlnet-preprocess@<type>>
  base + controlnet: <runware:101@1 + runware:25@1 (canny) | 29@1 (pose) | 27@1 (depth)>  (FLUX dev)
  steps / endStep: <~30 steps, endStep 10 for transform, 15+ for restyle>
  

  The control type drives every other line: it picks the preprocessor and the matching ControlNet, and how hard you hold structure (endStep) follows the intent (transform vs restyle).

• Prep the source for clean edges when using Canny: clear outlines, distinct features, decent contrast, minimal noise. Simplify a busy source before extracting.

Parameters that matter

• preProcessorType: canny, depth, openpose, seg, softedge. Must match the ControlNet you pair it with.
• lowThresholdCanny / highThresholdCanny: Canny sensitivity. Lower low = more edges, higher high = fewer edges.
• controlNet[].guideImage: the preprocessor output, not the raw source.
• controlNet[].weight: how strongly the map steers generation. 1.0 is a balanced start.
• controlNet[].startStep / endStep: the window the guide is active. 1/10 on ~30 steps is balanced. Raise endStep to hold structure harder, raise startStep to weaken it.
• Control mode: balanced / controlnet / prompt.
• Confirm every field name and range against the live schema (runware-run). Never guess.

Quality bar

• The output's composition (pose / outline / layout) matches the source while the intended look changed.
• Structure is held to the right degree: recognizable, not a rigid trace that blocks the new subject (raise endStep if too loose, lower it if too rigid).
• The map type fits the goal (Canny silhouette vs depth layout vs pose).
• For a set, every image reads as the same composition. Retry outliers by adjusting thresholds, weight, or the step window, not by re-rolling blindly.

Related skills

runware-run (call the two-step pipeline), runware-models (confirm the preprocessor + ControlNet are live), runware-prompting (write the new-look prompt). Sibling outcomes: character-consistency (hold a subject instead of a composition), product-photography (controlled product variations).