SwarmUI-KreaGoneWild

A SwarmUI extension with a small toolbox of Krea 2 tricks. Everything is gated on the selected model being detected as Krea 2 (krea-2 / krea-2/lora) and does nothing otherwise.

Three features, each backed by a third-party ComfyUI node pack and each living in its own nested sub-group under the Krea Gone Wild parameter group. Each sub-group's toggle is the feature's on/off switch — there is no separate enable param. Install buttons appear inside the sub-group of whichever node pack is missing.

1. KGW Rebalance

Applies ComfyUI-ConditioningKrea2Rebalance to both the positive and negative Krea 2 conditioning (and the refiner pass, if present).

• Toggle the KGW Rebalance group on to enable it.
• KGW Multiplier — overall conditioning multiplier.
• KGW per_layer_weights — comma-separated weights for the 12 Krea 2 conditioning layers.

2. KGW Ostris Edit

Runs ComfyUI-Krea2-Ostris-Edit edit LoRAs (ai-toolkit krea2 arch, model_kwargs.edit: true) — "Kontext-style" reference-image editing.

It wires the edit natively into SwarmUI rather than replacing SwarmUI's conditioning:

1. Your prompt image(s) are scaled to ~1 MP (÷16 snapped, matching the training preprocessing) and VAE-encoded, then attached as reference latents onto the positive and negative conditioning via the core ReferenceLatent node.
2. The model is wrapped in Krea2OstrisEditModelPatch so the Krea 2 DiT actually consumes those reference latents (index_timestep_zero).

Because SwarmUI's own SwarmClipTextEncodeAdvanced still builds the conditioning, all the advanced prompt syntax (<break>, weighting, [from:to:when], regions, …) keeps working on edits. Only the model patch comes from the ostris pack; the reference latents use core ComfyUI nodes — exactly the "chain conditioning from … Set Reference Latent nodes" path the pack documents.

• Toggle the KGW Ostris Edit group on to enable it. Requires a Krea 2 edit LoRA loaded and at least one prompt image. With no prompt image it is a no-op (the model patch passes through unchanged).
• KGW Edit Reference Megapixels — target size the reference is scaled to before VAE-encoding (default 1.0, matching ai-toolkit).

Note: the reference latents are attached to the main generation pass only; a separate refiner pass is not edit-patched.

3. KGW Krea2Edit

Runs comfyui-krea2edit — instruction-based editing with the krea2_edit identity LoRA. Different approach from the ostris pack: instead of reference latents on the conditioning, the source is prepended to the sampling sequence as a block of clean in-context tokens (RoPE frame=1), with the instruction encoded through Qwen3-VL alongside the image in the exact layout the LoRA trained on.

Two halves to the recipe. Only the model patch comes from the pack:

1. Grounding. SwarmUI already routes Krea 2 through SwarmClipTextEncodeAdvanced with the prompt images wired and its stock "krea2" template, which appends the vision blocks after the instruction with a Picture N: label. krea2_edit trained with them before the instruction and unlabelled. That node passes any non-magic llama_template string straight through to clip.tokenize and leaves the prompt text alone, so this extension just retemplates the existing positive and negative encoders in place. Result: the pack's grounded encode, with <break>, [from:to:when], [alter|nate], weighting and per-step scheduling all still working — none of which the pack's own Krea2EditGroundedEncode node supports.
2. In-context source. Krea2EditModelPatch wraps the model. It gets the raw prompt image(s) plus the VAE on the pixel path (immune to input/output resolution mismatches), the VAE-encoded source as the required source_latent, and the sampler's own latent as target_latent so the source encode happens before sampling instead of evicting the diffusion model mid-run.

Up to two prompt images feed the model patch: the first is the scene (ref_boost_a), the second the subject (ref_boost) — the training-matched order for the multi-reference LoRAs. Grounding sees every prompt image SwarmUI batched, one vision block each.

• Toggle the KGW Krea2Edit group on to enable it. Needs the krea2_edit LoRA and at least one prompt image; with no prompt image it is a no-op.
• KGW Krea2Edit Ref Boost — reference-fidelity dial on the last reference. 1.0 = off, above 1 pulls harder toward the reference's appearance. Optimal value is model-specific.
• KGW Krea2Edit Ref Boost A (advanced) — same dial for the first reference (the scene). No effect with a single prompt image.
• KGW Krea2Edit Fit Mode (advanced) — fit (training-matched, default) or crop (legacy) for older weights.
• KGW Krea2Edit Grounded Encode — on by default. Turn it off to keep SwarmUI's stock "krea2" template (still grounded, just with the training-mismatched vision-block placement).
• KGW Krea2Edit Grounding Pixels — caps the longest side of the images fed to the vision tower, via a core ImageScaleToMaxDimension node. 0 = leave SwarmUI's own prompt-image sizing alone. The LoRA trained with 384–768px jitter, so 640–768 is in-distribution.
• KGW Krea2Edit System Prompt (advanced) — replaces the system turn of the grounding template. Empty = krea2_edit's training default.

Note: the pack's own guidance — Turbo at 8 steps / CFG 1 for most edits, Raw at CFG 3 / 20 steps for removals, generate at ≤2 MP.

Usage

1. Install the required ComfyUI node(s) using the button in the matching sub-group, then restart the ComfyUI backend when prompted.
2. Select a Krea 2 model.
3. Toggle on the sub-group(s) you want (and, for either editing mode, add a prompt image + the matching edit LoRA).

The features are independent toggles. The two editing modes are alternative approaches to the same job — enable one, not both.