EditUtils: Flux2Klein Output Extractor lrzjason
Turn Klein's Encode Black Box Into Eight Usable Wires
- custom_output
- pad_info
- main_image
- vae_images
- ref_latents
- full_prompt
- llama_template
- no_refs_cond
- mask
When Flux2KleinEditTextEncode_EditUtils (or the general EditTextEncode_EditUtils) runs, it hands back a custom_output socket - an ANY-typed dictionary holding everything the encode produced. Everything is exactly the problem. Flux2KleinOutputExtractor_EditUtils splits that dictionary into eight typed outputs so you can actually wire the internals into the rest of your graph.
One input, custom_output, and eight outputs: pad_info, main_image, vae_images, ref_latents, full_prompt, llama_template, no_refs_cond, and mask. The extractor is mechanical - it pulls each key out of the dict and presents it - so the interesting part is knowing which outputs earn their keep.
The two you'll touch most: pad_info and main_image. pad_info is the dictionary describing the padding/cropping that happened when your reference was letterboxed to the 16-unit VAE grid. Hand pad_info and the encoded image to CropWithPadInfo_EditUtils on the way out and you undo the padding, getting back the original framing instead of black bars - this is the standard round-trip for "edit at a clean latent size, deliver at the original size." main_image is the processed main reference as an IMAGE, ready for a preview or a save.
The inspection crowd gets ref_latents and vae_images - the lists of reference latents and VAE-encoded images - which pair with ListExtractor_EditUtils and the Any2Image/Any2Latent converters when you want one item out as a typed wire. The diagnostics crowd gets full_prompt and llama_template, the assembled prompt string the encode built (useful for debugging why the model is ignoring a clause - read what actually went in). And no_refs_cond is the conditioning without reference latents attached, which is the one to reach for when you want the text encoding alone - though for that narrower job ClearRefLatents_EditUtils is the cleaner tool. mask just echoes the mask back out.
The changelog notes that the no_refs_cond output was recently fixed across the Output Extractors, so if yours comes back empty on an old install, update the pack before you debug anything else. Same install story as every node here: ComfyUI Manager (search "EditUtils") or git clone https://github.com/lrzjason/ComfyUI-EditUtils into custom_nodes, restart. It's pure plumbing - no models, no dependencies beyond stock ComfyUI.
Honest take: for a simple Klein edit you can ignore this node entirely; the encode node's conditioning and latent outputs cover you. Add the extractor when you need the round-trip crop, want to inspect refs, or are building a reusable graph where the intermediates deserve to be real wires.
Inputs (1)
| Name | Type | Default | Description |
|---|---|---|---|
| custom_output | ANY | — |
Outputs (8)
| Name | Type | Description |
|---|---|---|
| pad_info | ANY | — |
| main_image | IMAGE | — |
| vae_images | LIST | — |
| ref_latents | LIST | — |
| full_prompt | STRING | — |
| llama_template | STRING | — |
| no_refs_cond | CONDITIONING | — |
| mask | MASK | — |