RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2078868402176872450?inviteCode=rh-v1111

Krea 2 Advanced Cleanup - Single Image (Structure LoRA + Reverse Prompt)

Purpose: a single-image Krea 2 "cleanup" graph. It reverse-prompts the input image into a Krea 2 prompt, then rebuilds the image through depth-structure conditioning with a selectable number of render stages and a final 4x upscale. This is the single-image variant of the set.

Graph (as saved)

51 nodes / 89 links / 9 groups.

• Model: UNETLoader krea2_turbo_fp8.safetensors; CLIPLoader qwen3vl_4b_fp8_scaled.safetensors (type krea2); VAELoader qwen_image_HDR_vae_fp32_comfy.safetensors.
• Structure LoRA: one Krea2ControlLoRALoader on depth-control-lora.safetensors at strength 1.0 — a single shared structure LoRA, not a preset trio.
• Three depth stages: three Krea2ControlImageEncode → Krea2ControlApply pairs (first / middle / final size), all match_latent_size, lanczos, center, grayscale, per-image min/max.
• Sampling: three KSamplers at 8 steps / CFG 1 / euler / simple with denoise 0.55, 0.35 and 0.25; two LatentUpscale nodes (bislerp, 1024) saved disabled; final UpscaleModelLoader 4x-UltraSharp.pth + ImageUpscaleWithModel.
• Reverse prompt: Qwen3VLProcessor (open-source route) and RH_LLMAPI_Pro_Node (the closed route, model label GLM-4.6V) — the user console asks which source to use.
• User console: render stages (two or three), final long edge in pixels (2048/3072/4096, aspect preserved and 64-aligned), reverse-prompt source. ConditioningKrea2Rebalance is saved bypassed here; the ComfyUI-Krea2T-Enhancer patch is saved enabled.
• Output: SaveImage prefix KSK_Krea2_ReverseDepthHires.

How to use

Load the source image, pick two or three stages, set the final long edge, and choose the reverse-prompt route. Treat 2048 as the value-for-money point; the graph's own note says 3072/4096 increase VRAM and cost noticeably.

Limits and not-verified items

No inference was run for this listing: no speed, VRAM, fidelity or upscale-quality claim is made. The saved input image and notes are the author's demo content. Weights, LoRA, depth model, upscaler and the VLM routes are not bundled; check their licenses separately. The cover image is a promotional still, not a frame produced by this graph.

Tutorial: YouTube · Bilibili. Workflows supplied through AI-KSK; verify workflow, node, model, LoRA, depth model and cover rights separately. No blanket commercial or redistribution license is asserted.