Which models are the fastest?

If you want quick, guided image generation, fofr/latent-consistency-model is the fastest option — it can generate an image in under a second while supporting Canny edge detection and batch processing.

For simpler control tasks like sketch-to-image or QR pattern generation, jagilley/controlnet-scribble and andreasjansson/illusion also run quickly and work well for experimentation.

Which models offer the best balance of cost and quality?

black-forest-labs/flux-depth-pro and black-forest-labs/flux-canny-pro deliver professional-grade results with strong prompt adherence and high image quality.

They’re ideal when you need precise control (like retexturing or preserving shapes) without the heavy compute requirements of large multi-control setups.

What works best for keeping composition or structure consistent?

For structure-preserving edits, use black-forest-labs/flux-canny-pro. It locks in the edges and layout from your reference image so your generated version stays faithful to the original.

If you’re working with 3D shapes, architecture, or anything requiring spatial depth, black-forest-labs/flux-depth-pro preserves the scene’s geometry while allowing stylistic retexturing.

What works best for creative or experimental projects?

For artistic outputs, andreasjansson/illusion creates intricate patterns, spirals, and QR code art that still scans correctly.

jagilley/controlnet-scribble is another favorite — it turns simple line drawings into detailed, high-quality scenes, great for quick sketches or brainstorming visuals.

What’s the difference between edge, depth, and scribble control?

• Edge control (like black-forest-labs/flux-canny-pro or lllyasviel/controlnet-canny) focuses on the outlines and contours of your input, preserving structural details.
• Depth control (like black-forest-labs/flux-depth-pro or lllyasviel/controlnet-depth2img) maintains 3D spatial relationships, perfect for realistic lighting and object placement.
• Scribble or sketch control (like jagilley/controlnet-scribble) starts from rough shapes or doodles and fills in fine detail based on your prompt.

What kinds of outputs can I expect from these models?

All models output images that follow both your text prompt and structural guidance input.

Depending on the model, you might input a depth map, edge map, or sketch, and the output will match that composition while adopting the style you describe.

How can I self-host or push a model to Replicate?

Open-weight models like black-forest-labs/flux-depth-dev and black-forest-labs/flux-canny-dev can be self-hosted using Cog or Docker.

If you’re training or experimenting, you can fork an existing ControlNet model, define inputs in a replicate.yaml, and push it to your account to make it available on Replicate.

Can I use these models for commercial work?

Yes, most ControlNet-based models are cleared for commercial use, especially the black-forest-labs/flux-pro variants and fofr/latent-consistency-model.

Always confirm license terms on the model’s page — some community models or QR-based tools might have additional attribution requirements.

How do I use or run these models?

Upload a reference image or control map (like an edge or depth image), describe what you want in plain text, and run the model.

For example:

• “Turn this sketch into a watercolor painting” → jagilley/controlnet-scribble
• “Keep the building outline but change it to nighttime” → black-forest-labs/flux-canny-pro

You’ll get a generated image that follows both your guide and your description.

What should I know before running a job in this collection?

• Your input control image determines how closely the output follows structure—edges and depth maps give the most consistency.
• Lower creativity or guidance values produce more faithful edits, while higher values lead to freer, more artistic interpretations.
• Batch generation (available in fofr/latent-consistency-model) is efficient for large-scale workflows.

Any other collection-specific tips or considerations?

• Combine depth and edge control for highly detailed scene edits.
• If you’re doing architectural or interior work, adirik/interior-design can generate photorealistic rooms with ControlNet guidance.
• For stylized or surreal outputs, experiment with multiple control layers using pharmapsychotic/realvisxl-v3-multi-controlnet-lora.
• When generating artistic QR codes, test scanability with your phone to ensure the pattern still functions.