Nodes/Krea2_ComfyUI_Advanced/Eric Krea2 Reference Latents (Edit)
ComfyUI Node

Eric Krea2 Reference Latents (Edit)

A ComfyUI node in Eric/Krea2 with 6 inputs and 1 output.

By EricRollei·Created about a month ago·Updated 10 days ago· 5
Eric Krea2 Reference Latents (Edit)
  • krea2_pipeline
  • image1
  • image2
  • image3
  • ref_latents
max_ref_megapixels0.25
edit_recipeostris_t0
CategoryEric/Krea2

Inputs (6)

NameTypeDefaultDescription
krea2_pipelineKREA2_PIPELINE
image1IMAGEReference image 1 (VAE-encoded, appended to the image token sequence at t=0).
image2optIMAGEReference image 2 (optional).
image3optIMAGEReference image 3 (optional; published edit LoRAs are mostly trained for 1-2 refs).
max_ref_megapixelsoptFLOAT0.250.1–4Per-reference pixel cap before VAE encode (downscale only, never upscale). ai-toolkit edit training uses 1.0 MP, but community testing with the Style Reference LoRA found ~0.25 MP gives cleaner style transfer with fewer artifacts than the training-time default - raise it only if you need more structural detail from the reference. Note the Ultra node's ref_match_size then rescales the encoded reference again (in token space) to match its resolved Stage 1 size, so this cap mostly just bounds the encode-time VAE cost/detail ceiling, not the final grid size actually used. Each 1 MP ref adds ~2048 tokens to every denoise step, so 2-3 refs at 1 MP is slow.
edit_recipeoptCOMBOostris_t0Which trained reference mechanism to run - the LoRA and the recipe MUST match (a LoRA under the wrong recipe = the classic 'never as good as the author's examples' failure). ostris_t0: refs appended at their own grid with t=0 modulation - for style/reference LoRAs (ai-toolkit reference method, e.g. Krea 2 Style Reference). Pair with Vision Prompt template picture_n at ~0.15 MP. edit_frame: source prepended on the target grid, shared timestep - for identity/instruction edit LoRAs (ai-toolkit predict_velocity_edit, e.g. Krea 2 Identity Edit at strength 1.0). The Ultra node pixel-fits and re-encodes the source at its resolved Stage 1 size automatically (match your target aspect ratio to the source; trained <=2MP, so it conditions Stage 1 only in a multistage run). Pair with Vision Prompt template bare_edit at ~0.30-0.45 MP. Turbo/CFG-1 for most edits; removals need Raw at CFG ~3 with an EMPTY negative prompt.

Outputs (1)

NameTypeDescription
ref_latentsKREA2_REF_LATENTS