How the three Krea2 generation pipelines divide the work
Photoreal generation, character casting, and a three-stage upscale pass — what each pipeline handles and why splitting them beats one universal form.
Why three pipelines instead of one
Most AI image tools take the same approach: one big form, every parameter exposed. Past a certain count the interface becomes a control panel and users no longer know what to adjust.
Krea2 splits into three independent pipelines, each exposing only what it genuinely needs:
| Pipeline | What it does | Parameters you touch |
|---|---|---|
| Photoreal generation | Prompt to high-realism photo | 5–6 (subject, scene, lighting, aspect, sampling strength) |
| Character casting | Consistent character across images | 1 reference + 4 (expression, pose, outfit, scene) |
| Three-stage upscale | Upscale to 2K/4K with detail recovery | Target size + 3 refinement strengths |
The benefit of splitting: each pipeline’s defaults are tuned for its own task. Photoreal generation and casting do not want the same CFG. Merged into one form, neither gets good defaults.
Pipeline one: photoreal generation
The core is two-pass sampling.
Ordinary sampling walks from noise to final image in one go, which tends to over-smooth fine detail. Two-pass splits it: the first pass covers most of the journey and locks composition, the second re-noises from the midpoint and re-walks to the end, specifically recovering skin texture, hair strands, and fabric weave.
The cost is roughly double the time. So the default puts it on a “high detail” preset, while quick previews use single pass.
Use two-pass when: delivering a final asset, printing, or upscaling to 4K. Skip it when: exploring composition or iterating on prompts.
Pipeline two: character casting
This solves “the same person looks different in every image”.
The approach is not plain img2img. It decomposes the reference into separately injected dimensions:
- Facial structure — proportions and face shape
- Skin tone and texture — the first thing to drift; plain img2img averages it away
- Hair — controlled on its own layer
- Style — photoreal / film / illustration
Multiple reference slots let you feed front, profile and full-body shots at once; the model synthesizes a more stable character identity.
Measured experience: 3 references clearly beat 1, but returns diminish past 5 and VRAM gets tight.
Pipeline three: three-stage upscale
Upscaling is not interpolation. A straight 4× resize gives you a blurry big image.
The three stages:
- Stage one — 2× upscale plus light re-noising, recovering macro structure
- Stage two — another 2× plus moderate repaint, recovering mid-frequency detail
- Stage three — no upscaling, only low-strength refinement, unifying texture and ironing out inconsistencies
Stage three is the one people skip, and it contributes the most to perceived quality. Stages one and two inevitably repaint different regions at different strengths; stage three levels that out.
Why not just use the ComfyUI canvas
ComfyUI is powerful and repetitive. Twenty variants of the same shot means rewiring the graph and resetting parameters each time.
My approach wraps the graph into a form. Parameter changes land in an “injector” that mutates an in-memory copy of the workflow — the original workflow file on disk is never touched. You keep ComfyUI’s capability without staring at a node graph daily.
Choosing among 53 LoRAs
The built-in gallery ships 53 LoRAs with 529 example images and prompts.
How to pick:
- Look at the images, not the names. A LoRA called “cinematic” might just darken everything.
- Start at 0.6 strength. Full 1.0 frequently breaks anatomy in the base model.
- Two LoRAs stacked: keep total strength under 1.2.
FAQ
- Does it need internet?
- No. The kernel ships with the installer; once models are downloaded everything runs offline. No request ever uploads your images.
- What GPU is needed?
- 8 GB VRAM runs base generation at 1024×1024. Character casting wants 12 GB+. Three-stage upscale to 4K wants 16 GB+.
- Can I import my own workflows?
- Yes. The workflow directory reads standard ComfyUI JSON. But the parameter form is only mapped for the three built-in pipelines; imported workflows fall back to a generic form.