Krea 2's cleanup graph is not one workflow with three buttons. It is one recipe — three depth-conditioned passes over a RAW high-noise stage and Turbo mid/low stages — wrapped in three different I/O shells. This episode ships all three: a local folder batch, a RunningHub online batch, and a single-image version.
No inference was run for this release. Everything below is read from the saved graphs: node counts, sampler values, LoRA strengths and the preset logic. There is no timing, no VRAM number and no quality comparison here.
| Variant | File | Nodes / links | I/O shell |
|---|---|---|---|
| Local batch | 01 | 92 / 154 | folder loop + batch |
| Online batch | 02 | 104 / 169 | RunningHub ZIP / batch / LLM nodes |
| Single image | 03 | 89 / 149 | one LoadImage |
The shared spine
All three files carry the same core, and it is worth naming precisely because the file titles do not.
Three depth stages. Each file has three Krea2ControlImageEncode nodes — one for the high-noise stage, one for the mid stage, one for the final size — each feeding its own Krea2ControlApply. Each stage also gets a Krea2ControlLoRALoader on depth-control-lora.safetensors, saved at strengths 1.00, 0.85 and 0.65. Those are three stages of one preset-driven chain, not three separate workflows.
Then there are three more LoraLoaderModelOnly nodes pointing at the same LoRA file with strength 0. Those are empty slots. If you read the node count and conclude "six LoRAs are active", you will be wrong.
Two base models. krea2_raw runs the high-noise shaping stage (the online variant loads krea2_raw_bf16_0630 instead), and krea2_turbo_fp8 runs the mid and low stages. Text encoding is qwen3vl_4b_fp8_scaled.safetensors with type krea2; the VAE is qwen_image_HDR_vae_fp32_comfy.safetensors.
Depth is generated in-graph. DA3 nodes (LoadDA3Model, DA3Inference, DA3Render) produce the depth maps; the encodes use match-latent-size, lanczos resize, center crop, grayscale and per-image min/max.
Sampling values, as saved. Three SamplerCustomAdvanced nodes share one SamplerEulerAncestral (1, 1). The shared RAW/Turbo schedule is 36 steps at denoise 0.92; the Turbo upscale refine is 10 steps at denoise 0.35. The active intermediate KSampler is 10 steps, CFG 1, euler_ancestral, simple, denoise 0.40. The legacy KSamplers — 8 steps at 0.95 or 0.90, 10 steps at 0.30, 8 steps at 0.34 — are all saved bypassed, so do not quote them as the current setup.
Upscale. 4x-UltraSharp through UpscaleModelLoader + ImageUpscaleWithModel, plus two LatentUpscale nodes.
The preset dial. One integer picks the behaviour: 0 is tightly controlled and faithful, 1 is balanced (the default), 2 is a diverging high-noise repaint. Switching it does not change one number — it moves the three denoise values, the Split checkpoints, the RAW CFG and the depth-LoRA strengths together. In preset 2 the depth LoRA strength is set to 0. That is why "just bump the denoise" is not the same thing as changing the preset.
01 — Local batch: the folder loop
The local variant adds a folder loop: easy loadImagesForLoop feeds easy forLoopEnd, with easy batchAnything handling the batch. Output prefix is Krea2_本地批量洗图_2.4_RAWTurbo.
Two saved-state details matter. ConditioningKrea2Rebalance is active here (it is bypassed in the single-image file), and the ComfyUI-Krea2T-Enhancer patch plus the old 2048×2048 scale node are saved bypassed. The graph is configured to run clean without the enhancer.
02 — Online batch: RunningHub nodes take over I/O
The online variant swaps the local loop for platform nodes: RHUploadZip for a ZIP of images, RHBatchImages: and RH_ExtractImage for batch extraction, and RH_LLMAPI_Pro_Node for the caption/JSON path. The RAW stage loads the krea2_raw_bf16_0630 weight here.
If you run this file outside RunningHub, those nodes will not resolve. That is not a broken workflow — it is a platform-bound workflow. Either run it on RunningHub or replace the RH nodes with local loaders.
03 — Single image: the cleanest read of the core
The single-image file is the best place to study the core because there is no batch plumbing: one LoadImage, the same three depth stages, the same preset dial, and ConditioningKrea2Rebalance saved bypassed. Output prefix is Krea2_单图洗图_2.4_RAWTurbo.
How I would approach a first run
Start with the single-image file at preset 1, and look at the three depth stages before touching anything. If the structure drifts, go to preset 0 before you start editing sampler values; the preset is doing more than one thing at a time. If you want a more aggressive re-render, preset 2 explicitly zeroes the depth LoRA — that is the intended "let it repaint" mode, not a bug.
For batch work, use the local variant unless you specifically need the RunningHub LLM caption path. The local loop is simpler to debug and does not depend on platform nodes.
What this is not
It is not a benchmark, and it is not a commercial-license grant. The video's framing about a "member-grade commercial workflow" is the author's description; the files themselves carry no license grant for the weights, the LoRA, the depth model or the example images. Check those separately.
It is also not an installation guide: the graphs reference custom nodes (DA3 nodes, Krea2 control nodes, easy loop nodes, and RunningHub nodes in the online variant) that must exist at compatible versions.
Getting the workflows
Published by this channel. The graphs are free; weights, LoRA, depth model and example media are not bundled.
- Hugging Face: AI-KSK/krea2-cleanup-three-presets-284 — three JSONs, source covers, SHA256 manifest. Start here.
- ModelScope (魔搭): Sunjian520/krea2-cleanup-three-presets-284
- Civitai: 01 local batch · 02 online batch · 03 single image
-
Run without a local GPU: 01 local batch · 02 online batch · 03 single image — these carry the author's invite code
rh-v1111. - Tutorial: YouTube · Bilibili
Verify workflow, node, model, LoRA, input-image and cover rights separately. No blanket commercial or redistribution license is asserted for the graphs or the example media.
Top comments (0)