Flux Kontext Dev Two-Image Edit with NAG

How it works

The workflow loads flux1-kontext-dev.safetensors with UNETLoader, plus clip_l, t5xxl_fp16 and ae.safetensors. It takes two images, resizes each with ImageResizeKJv2 and FluxKontextImageScale, and can remove their backgrounds with BiRefNetRMBG set by the remove_background toggle. ImageStitch places them next to each other, and VAEEncode with ReferenceLatent hands both to the model.

A Preference_Latent toggle chooses between an empty latent and the first concept as the starting point. The SwitchAny nodes route these choices.

Prompting can be automatic. With enhance_prompt on, a system prompt task generator and an OpenAI provider node write a Kontext-style prompt from your instruction. Sampling uses KSamplerWithNAG with FluxGuidance, so your negative prompt has an effect. RepeatLatentBatch creates batches, and an ImageComparer shows the result.

Tips

• Turn on remove_background when you want effects to stay inside the character outline.
• Toggle Preference_Latent to compare starting from an empty latent versus the first concept.
• Add terms like short legs or big head to the negative prompt if you see those problems.
• If no OpenAI key is on the machine, turn enhance_prompt off and write the prompt yourself.

What you can use it for

• Blend two characters into a new design
• Merge design elements from two references
• Test negative prompts in Kontext edits
• Concept art iterations