IAMCCS Krea 2 Identity Multi-Gen

Krea 2, if you missed it, is the 12B diffusion transformer that took the open-weights world by storm in mid-2026 - a genuinely undistilled base model, permissively licensed, with a notorious quirk: the open weights went through an alignment pass the API version didn't, so identity preservation in particular wants a careful hand. This node is IAMCCS's answer to the "one identity, many shots" problem: paste a newline-separated list of prompts and it runs a full Krea 2 image-edit generation for each one, from a single reference image, with identity-preservation baked into the conditioning at every step.

The mechanism is worth understanding because it's both clever and demanding. It doesn't reinvent the Krea 2 pipeline - it composes it at runtime. The node calls Krea2EditModelPatch to patch your model with the source latent (and an optional image_b for a second reference), then Krea2EditGroundedEncode for grounded conditioning, ConditioningKrea2Rebalance with your per_layer_weights, and samples with ClownsharKSampler_Beta (the RES4LYF beta sampler). Then it loops: one encode+sample+decode per prompt line, stacking all results into one batched images output.

That means the install story is the headline gotcha: this node requires the Krea 2 nodes to already be installed. If Krea2EditModelPatch, Krea2EditGroundedEncode, or ConditioningKrea2Rebalance aren't registered, the node fails at runtime with a clear "install/update the Krea 2 Edit, Conditioning Rebalance and IAMCCS custom nodes" error. Same for ClownsharKSampler_Beta. It will not silently skip - and that's the right behavior, because a multi-gen run that silently drops its identity machinery would waste your time.

The inputs that matter:

• image (and optional image_b) - your identity reference(s). The system_prompt default is a genuinely good identity brief ("Prioritize facial geometry, apparent adult age, eye color, hair characteristics...") - it rides along on every encode and it's the main reason generations stay on-identity.
• multi_prompt + separator - the list (newline by default) and the splitter. Blank lines are skipped; an empty list returns a zero-count report instead of erroring.
• seed + seed_mode - fixed (same seed every prompt), increment (seed + index), or random. Increment is the one you want for variation batches.
• ref_boost (1.15) / ref_boost_a (1.0) - reference conditioning strength knobs that get passed straight into the Krea 2 model patch. Push ref_boost when faces drift.
• grounding_px (768), per_layer_weights (the long default string, with heavy weights late in the list), rebalance_multiplier (4) - the grounded-encode and rebalance controls. The defaults are tuned; change them deliberately.
• width/height (1024), steps (8 default - Krea 2 is a few-step model), cfg (1.0), denoise (1.0), eta, sampler_name (exponential/ddim), scheduler (beta57).

Outputs: images (a batch of all generations, decode-flattened), count (how many ran), and report (JSON with the actual seeds used and how many references were active - keep this, it's your reproducibility record).

Installation: search IAMCCS in ComfyUI Manager, or clone https://github.com/IAMCCS/IAMCCS-nodes.git into custom_nodes, restart - then install the Krea 2 edit/conditioning packs and the RES4LYF pack for the beta sampler, because this node composes them. Model side: a Krea 2 checkpoint and the Qwen-Image VAE (Krea 2's VAE family).

The trap: rebalancing at default rebalance_multiplier 4 with per_layer_weights set to identity-heavy late layers is what holds the face together - if your batch drifts, it's almost always those two knobs plus ref_boost, not the prompts.