ausboss
ec50d3acfe
Krea 2 accepts reference latents, but they arrive with no position, so the model treats them as a loose style hint: it borrows the look and reinvents the content. That is the whole reason an outpaint drifts instead of continuing the photo. Krea 2 Encode attaches the reference latents and emits the negative from the same node, so a turbo graph at CFG 1.0 stops carrying a second text encode that does nothing. Krea 2 Outpaint Model Patch registers those reference tokens into the target grid at the rectangle the stitcher records, on frame axis 1. Load Image + Pad gained a `reference` output - the unpadded source, fitted to a 384px long edge and a multiple of 16, because the VAE downsamples by 8 and the DiT patchifies by 2. It is APPENDED after `stitcher`, never inserted: a workflow stores links by output slot index, so any other position silently rewires every saved graph. The stitcher now records where the source sits, as source_bbox in pixels and bbox_normalized in 0..1. Padding knows the rectangle exactly, so it rides with the canvas instead of a parallel wire that can be left unplugged. apply_stitch reads neither key, so an older stitcher stitches identically. Ported from the lab rather than copied: the lab's _krea2_helpers had its own pad_image, feather_pad_mask, plan_pad_canvas, parse_fill_color and build_canvas_stitcher. Duplicating the padding engine here would let the outpaint path and the inpaint path disagree about what a stitcher is, so the public module keeps only what is specific to Krea 2 and uses the existing helpers for the rest. Two departures from the lab version. The model patch declared a kv_cache input and never read it - the cache was always on - so it is wired up now. And it imports comfy's flux internals when you run it rather than at startup: a core release that moves one then surfaces as an error on that node, instead of deleting it from the menu and leaving saved workflows reporting it missing.
44 lines
2.1 KiB
Markdown
44 lines
2.1 KiB
Markdown
# Krea 2 Encode
|
|
|
|
Encodes a Krea 2 prompt pair and attaches reference latents, in one node.
|
|
|
|
Wire a VAE and a reference image and the reference is encoded and appended to
|
|
the positive conditioning. Leave them unwired and this is a plain two-prompt
|
|
encoder. The negative comes out of the same node, so a turbo graph running at
|
|
CFG 1.0 still has something to plug in without a second text encode sitting
|
|
there doing nothing.
|
|
|
|
## Controls
|
|
|
|
- **clip**: The Krea 2 text encoder.
|
|
- **prompt**: Describe the whole finished canvas, not just the new area — the
|
|
model generates all of it and matches the reference where it must.
|
|
- **negative_prompt**: Ignored at CFG 1.0, which is where turbo runs. Empty is
|
|
the normal case there.
|
|
- **vae**: Needed to turn the reference images into latents. Without it the
|
|
references are skipped and only the prompts are encoded.
|
|
- **reference**: The `reference` output of Load Image + Pad 🆎 — the unpadded
|
|
source. Any image works; it is fitted to a multiple of 16 first.
|
|
- **extra_image**: A second reference, e.g. a style or character plate.
|
|
- **vlm_reference**: Also show the references to the vision tower, so the text
|
|
encoder describes them. Off for outpaint: the latents already carry the
|
|
picture, and the description tends to pull the result toward a paraphrase of
|
|
the source.
|
|
|
|
## Outputs
|
|
|
|
- **positive**: Prompt conditioning with the reference latents attached.
|
|
- **negative**: Negative prompt conditioning.
|
|
|
|
## Notes and limitations
|
|
|
|
- **References are downscaled to a 384px long edge and snapped to a multiple
|
|
of 16.** The VAE downsamples by 8 and the DiT patchifies by 2, so an edge off
|
|
a multiple of 16 lands on a partial patch and the token grid stops lining up
|
|
with the canvas grid. The cap is deliberate — a reference is there to say
|
|
"this is the picture", and the canvas latent carries the detail.
|
|
- Attaching reference latents alone does not tell the model *where* they go.
|
|
Pair this with Krea 2 Outpaint Model Patch 🆎 for outpainting.
|
|
- On a core without the Krea 2 prompt template, `vlm_reference` still works —
|
|
it falls back to tokenizing without the template.
|