🎯 Edit Encoder

If you've ever hand-built an edit workflow for a FLUX.2-family model, you know the drill: for every reference image you chain ImageScaleToTotalPixels → VAEEncode → ReferenceLatent, plus a separate CLIPTextEncode for the prompt, and you keep all those wires straight across ten images. FiL Edit Encoder replaces that whole scattered chain with one node: prompt plus as many reference images as you can feed it in, one CONDITIONING out, ready for the sampler. It targets the instruction-edit family that runs on vision-language text encoders - Krea 2, Klein, FLUX.2 Dev, and the Kontext-lineage models that read reference_latents the FLUX.2 way.

This family edits the way it does because the text encoder is a vision-language model (Qwen3-VL on Krea 2, Mistral-3 on FLUX.2) - a "CLIP" node here isn't just reading your words, it can look at pictures. That's the whole game.

Two channels, one decision

The most important knob on this node is reference_mode, because it picks which half of the model receives your references, and they produce visibly different results:

• vision (default) - the reference goes to the text encoder, which looks at it without putting anything into the frame. Nothing of your source image gets baked into the canvas. This is Krea 2's native behaviour and the mode the node is aimed at.
• latents - the reference is VAE-encoded and concatenated into the frame's own tokens, which is the FLUX.2/Kontext behaviour. This is what tiles the source into the output - you're re-denoising on top of the picture.
• both - only sensible if the model was trained for it; measured on Krea 2 it edits correctly but shows the reference bleeding through as a mosaic.

The trap people fall into: they expect Kontext-style "paste my photo in and repaint it" and leave it on vision, then wonder why the output ignores the source layout - or the reverse, they get their source tiled into the frame when they only wanted it read. Decide which you want first, then set the mode.

The inputs and outputs that matter

For a first working graph you set four things: clip (the edit model's text encoder - wire the CLIP socket from your model's loader), prompt (the edit instruction: "keep the pose, change her dress to red"), images (auto-growing reference slots), and reference_mode. vae is only required when reference_mode uses latents or when you wire a mask - if you forget it, the node raises a clear error telling you exactly that, rather than failing mysteriously in the sampler.

Outputs: conditioning feeds the sampler's positive input. latent is the first reference VAE-encoded - the thing you sample on top of, carrying your mask aligned when one is wired. references shows the prepared copies the model actually received (wire a Preview Image to see what it read - great for debugging). summary is plain text describing what the run did with each reference, and it flags settings that look like bugs but aren't.

The deeper controls - per-reference reference_cards as JSON roles ([{"role": "style"}] washes that picture to a colour field so there's no subject left to copy), vision_megapixels (0.15 is core's value; push toward 1 MP to hold structure and detail at the cost of tokens), and prompt_strength (cross-fades how loudly the written instruction speaks against the pictures) - are for when the default "use it as-is" isn't cutting it. Each reference you add grows VRAM and sampling time, so the auto-growing image list isn't a licence to dump your whole folder into it.

Installing FiL_Design_ImageMind

The pack needs ComfyUI 0.3.60+ (V3 node API). ComfyUI Manager → search FiL_Design_ImageMind → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/FiL-Design-Ai/FiL_Design_ImageMind.git
pip install -r FiL_Design_ImageMind/requirements.txt

Windows portable installs: run install_requirements.bat. Dependencies are light (requests, aiohttp, PyYAML, Pillow, numpy, pydantic), no model downloads, no API keys - this node is pure conditioning work and doesn't touch the pack's LLM-provider side at all. The nodes appear under the 🎨 FiL Design menu.

Troubleshooting

• A clear error about vae - you're in latents/both mode or you wired a mask without the VAE. Wire the edit model's VAE (or switch to vision).
• The source image is baked into the result - you're in latents; that's what it does. Drop to vision if you only wanted the model to study it.
• Mosaic/bleed-through with both on Krea 2 - expected and measured; use vision.
• Reference looks too loosely held - raise vision_megapixels toward 1.0 MP; the default 0.15 keeps the subject only loosely by design.
• Nothing changed or something odd happened - read the summary output and preview references; the node tells you what it actually did with each picture, and it's almost always a mode misunderstanding rather than a bug.

The pack is young, small and MIT-licensed - file issues upstream if a run misbehaves.