https://youtu.be/aCEuir7zy-w 
 
This workflow is the local batch restyle version of the Krea 2 RAW + Turbo refinement setup. It is designed for creators who want to process many source images on a local ComfyUI environment while preserving the original subject identity, pose, composition, lighting logic, color palette, and depth relationships. The goal is not to redraw everything from zero. It is a controlled image-to-image restoration and restyle pipeline for cleaning edges, improving material texture, increasing visual polish, and making a batch of images feel more consistent. 
 
The inspected graph uses two active Krea 2 UNET routes: `krea2_raw.safetensors` for the high-noise shaping stage and `krea2_turbo_fp8.safetensors` for the lower-noise convergence and refinement stages. The text route uses `qwen3vl_4b_fp8_scaled.safetensors` as the Krea2 CLIP encoder, and decoding is handled by `qwen_image_HDR_vae_fp32_comfy.safetensors`. Depth control is active in three stages through `depth-control-lora.safetensors`, with strengths 1.00, 0.85, and 0.65, so the workflow can keep source structure while still allowing the RAW stage to rebuild detail. 
 
The local batch version is built around a manual JSON prompt block, making it easier to keep a stable instruction across a folder of images. The active chain includes a 36-step RAW/Turbo sigma schedule, RAW high-noise shaping, Turbo source-resolution convergence, intermediate and final latent upscales to 1024 x 1024, a 10-step Turbo refinement pass, and an active 4x-UltraSharp image upscale route. Bypassed legacy KSampler and disconnected final scaling nodes are not treated as active features here. 
 
Main features: 
 
- Local batch Krea 2 restyle workflow 
- RAW high-noise shaping with Krea 2 Raw 
- Turbo low-noise refinement with Krea 2 Turbo FP8 
- Qwen3VL 4B FP8 Krea2 text encoder 
- Qwen Image HDR VAE route 
- Three-stage Krea2 depth control 
- Depth ControlNet LoRA strengths 1.00, 0.85, and 0.65 
- Manual JSON prompt block for batch consistency 
- 1024 x 1024 latent refinement stages 
- Active 4x-UltraSharp image upscale route 
- Best for local batch cleanup, restoration, and restyle tasks 
 
Suggested workflow: 
 
Prepare a folder of source images with similar quality targets, keep the JSON prompt direct, and use the source image as the authority for composition, identity, pose, lighting, and color. Start with the default prompt before adding stronger style language. If the output drifts too far from the original image, reduce creative wording and keep depth control strong. If the output is too conservative, loosen the prompt and let the RAW stage rebuild more detail. 
 
Related resources: 
 
Krea 2 Model Collection (Quark): https://pan.quark.cn/s/07bdc81784ce 
Krea 2 Raw: https://huggingface.co/krea/Krea-2-Raw 
Krea 2 Turbo: https://huggingface.co/krea/Krea-2-Turbo 
Krea 2 Depth ControlNet LoRA: https://huggingface.co/Patil/Krea-2-depth-controlnet 
 
RunningHub Workflow 
 
Try the workflow online right now - no installation required. 
Workflow: https://www.runninghub.ai/post/2080681468573085697?inviteCode=rh-v1111 
 
If the results meet your expectations, you can later deploy it locally for customization. 
 
Fan Benefits: Register to get 1000 points + daily login 100 points - enjoy 4090 performance and 48 GB super power. 
 
Bilibili Updates (Mainland China &amp; Asia-Pacific) 
 
If you are in the Asia-Pacific region, you can watch the video below to see the workflow demonstration and creative breakdown. 
Bilibili Video: https://www.bilibili.com/video/BV1GBgi6JEwL/ 
 
Support Me on Ko-fi 
 
If you find my content helpful and want to support future creations, you can buy me a coffee. 
Every bit of support helps me keep creating. 
Ko-fi: https://ko-fi.com/aiksk 
 
Business Contact 
 
For collaboration or inquiries, please contact aiksk95 on WeChat. 
 
打开下方链接即可在线体验，无需安装。 
工作流：https://www.runninghub.ai/post/2080681468573085697?inviteCode=rh-v1111 
 
如果觉得效果理想，也可以在本地进行自定义部署。 
 
粉丝福利：注册即送 1000 积分，每日登录 100 积分，可体验 4090 / 48G 级别算力。 
 
Bilibili 更新（中国大陆及亚太地区） 
 
如果你在中国大陆或亚太地区，可以通过下方视频查看该工作流的实测效果与构思讲解。 
B站视频：https://www.bilibili.com/video/BV1GBgi6JEwL/ 
 
本期 Krea 2 相关模型资源： 
Quark: https://pan.quark.cn/s/07bdc81784ce 
Krea 2 Raw: https://huggingface.co/krea/Krea-2-Raw 
Krea 2 Turbo: https://huggingface.co/krea/Krea-2-Turbo 
Krea 2 Depth ControlNet LoRA: https://huggingface.co/Patil/Krea-2-depth-controlnet 
 
这些资源主要面向本地用户，方便进行创作与学习。

Krea2 RAW-Turbo Local Batch Restyle Workflow

krea2RAWTurboLocal_v10.json

v1.0

<div data-youtube-video="https://youtu.be/aCEuir7zy-w"><iframe src="https://www.youtube.com/embed/aCEuir7zy-w" width="853" height="480" allowfullscreen></iframe></div> 
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2080681468573085697?inviteCode=rh-v1111
<h2>Krea 2 Cleanup 2.6 - Local Batch (Three-Stage Depth Control + Presets)</h2>
Purpose: a batch image "cleanup" graph that walks a folder of images through three depth-conditioned stages (high-noise shaping, mid-noise structure hold, low-sigma detail convergence), with RAW and Turbo base models and a 4x upscale pass at the end. This is the local-folder variant of the set.
<h2>Graph (as saved)</h2>
92 nodes / 154 links / 39 groups.
<ul>
<li>Depth chain: three <code>Krea2ControlImageEncode</code> stages (high / mid / final size), each feeding its own <code>Krea2ControlApply</code> and its own <code>Krea2ControlLoRALoader</code> on <code>depth-control-lora.safetensors</code> at strengths 1.00 / 0.85 / 0.65 (preset-driven).</li>
<li>Three further <code>LoraLoaderModelOnly</code> slots also point at <code>depth-control-lora.safetensors</code> but are saved at strength 0 - they are empty / disabled slots, not active LoRAs.</li>
<li>Models: <code>krea2_raw.safetensors</code> for the high-noise first stage, <code>krea2_turbo_fp8.safetensors</code> for the mid and low stages. CLIP <code>qwen3vl_4b_fp8_scaled.safetensors</code> (type <code>krea2</code>), VAE <code>qwen_image_HDR_vae_fp32_comfy.safetensors</code>.</li>
<li>Depth maps are generated in-graph with the DA3 nodes (<code>LoadDA3Model</code> / <code>DA3Inference</code> / <code>DA3Render</code>); the three encodes use <code>match_latent_size</code>, lanczos, center, grayscale, per-image min/max.</li>
<li>Sampler chain: three <code>SamplerCustomAdvanced</code> nodes sharing one <code>SamplerEulerAncestral</code> (1, 1). Schedulers: shared RAW/Turbo schedule 36 steps denoise 0.92; Turbo upscale refine 10 steps denoise 0.35. Active intermediate KSampler: 10 steps, CFG 1, euler_ancestral, simple, denoise 0.40. Legacy KSamplers (8 steps 0.95, 10 steps 0.3, 8 steps 0.34) are saved bypassed.</li>
<li>Upscale: <code>UpscaleModelLoader</code> 4x-UltraSharp + <code>ImageUpscaleWithModel</code>; two <code>LatentUpscale</code> nodes.</li>
<li>Preset dial: one integer selects 0 = tightly controlled / faithful, 1 = balanced (default), 2 = diverging / high-noise repaint; it switches the three denoise values, Split checkpoints, RAW CFG and depth-LoRA strengths together (preset 2 sets the depth LoRA to 0).</li>
<li>Batch loop: <code>easy loadImagesForLoop</code> + <code>easy forLoopEnd</code> + <code>easy batchAnything</code>; SaveImage prefix <code>Krea2_本地批量洗图_2.4_RAWTurbo</code>.</li>
</ul>
<h2>How to use</h2>
Point the loader at a folder, choose the preset number, and leave the depth-LoRA trio alone unless you deliberately override the preset. The example prompt path uses a JSON/LLM caption (RH_LLMAPI-style node present in the online variant); this local variant expects the prompt to be supplied from the JSON field or the CLIP text node.
<h2>Limits and not-verified items</h2>
No inference was run for this listing: no speed, VRAM, batch time, fidelity or upscale-quality claim is made. The saved nodes and notes are the author's configuration, not measured results. Weights, LoRA, depth model and input media are not bundled; check their licenses separately. The cover image is a promotional still from the source video, not a frame produced by this graph.
Tutorial: <a href="https://www.youtube.com/watch?v=aCEuir7zy-w" rel="ugc">YouTube</a> · <a href="https://www.bilibili.com/video/BV1GBgi6JEwL" rel="ugc">Bilibili</a>. Workflows supplied through AI-KSK; verify workflow, node, model, LoRA, input-image and cover rights separately. No blanket commercial or redistribution license is asserted.

Krea 2 Cleanup 2.6 - Local Batch (Three-Stage Depth Control + Presets)

<div data-youtube-video="https://youtu.be/aCEuir7zy-w"></div> 
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2080681468573085697?inviteCode=rh-v1111
Krea 2 Cleanup 2.6 - Local Batch (Three-Stage Depth Control + Presets)
Purpose: a batch image "cleanup" graph that walks a folder of images through three depth-conditioned stages (high-noise shaping, mid-noise structure hold, low-sigma detail convergence), with RAW and Turbo base models and a 4x upscale pass at the end. This is the local-folder variant of the set.
Graph (as saved)
92 nodes / 154 links / 39 groups.
<ul>
<li>Depth chain: three <code>Krea2ControlImageEncode</code> stages (high / mid / final size), each feeding its own <code>Krea2ControlApply</code> and its own <code>Krea2ControlLoRALoader</code> on <code>depth-control-lora.safetensors</code> at strengths 1.00 / 0.85 / 0.65 (preset-driven).</li>
<li>Three further <code>LoraLoaderModelOnly</code> slots also point at <code>depth-control-lora.safetensors</code> but are saved at strength 0 - they are empty / disabled slots, not active LoRAs.</li>
<li>Models: <code>krea2_raw.safetensors</code> for the high-noise first stage, <code>krea2_turbo_fp8.safetensors</code> for the mid and low stages. CLIP <code>qwen3vl_4b_fp8_scaled.safetensors</code> (type <code>krea2</code>), VAE <code>qwen_image_HDR_vae_fp32_comfy.safetensors</code>.</li>
<li>Depth maps are generated in-graph with the DA3 nodes (<code>LoadDA3Model</code> / <code>DA3Inference</code> / <code>DA3Render</code>); the three encodes use <code>match_latent_size</code>, lanczos, center, grayscale, per-image min/max.</li>
<li>Sampler chain: three <code>SamplerCustomAdvanced</code> nodes sharing one <code>SamplerEulerAncestral</code> (1, 1). Schedulers: shared RAW/Turbo schedule 36 steps denoise 0.92; Turbo upscale refine 10 steps denoise 0.35. Active intermediate KSampler: 10 steps, CFG 1, euler_ancestral, simple, denoise 0.40. Legacy KSamplers (8 steps 0.95, 10 steps 0.3, 8 steps 0.34) are saved bypassed.</li>
<li>Upscale: <code>UpscaleModelLoader</code> 4x-UltraSharp + <code>ImageUpscaleWithModel</code>; two <code>LatentUpscale</code> nodes.</li>
<li>Preset dial: one integer selects 0 = tightly controlled / faithful, 1 = balanced (default), 2 = diverging / high-noise repaint; it switches the three denoise values, Split checkpoints, RAW CFG and depth-LoRA strengths together (preset 2 sets the depth LoRA to 0).</li>
<li>Batch loop: <code>easy loadImagesForLoop</code> + <code>easy forLoopEnd</code> + <code>easy batchAnything</code>; SaveImage prefix <code>Krea2_本地批量洗图_2.4_RAWTurbo</code>.</li>
</ul>
How to use
Point the loader at a folder, choose the preset number, and leave the depth-LoRA trio alone unless you deliberately override the preset. The example prompt path uses a JSON/LLM caption (RH_LLMAPI-style node present in the online variant); this local variant expects the prompt to be supplied from the JSON field or the CLIP text node.
Limits and not-verified items
No inference was run for this listing: no speed, VRAM, batch time, fidelity or upscale-quality claim is made. The saved nodes and notes are the author's configuration, not measured results. Weights, LoRA, depth model and input media are not bundled; check their licenses separately. The cover image is a promotional still from the source video, not a frame produced by this graph.
Tutorial: <a href="https://www.youtube.com/watch?v=aCEuir7zy-w" rel="ugc">YouTube</a> · <a href="https://www.bilibili.com/video/BV1GBgi6JEwL" rel="ugc">Bilibili</a>. Workflows supplied through AI-KSK; verify workflow, node, model, LoRA, input-image and cover rights separately. No blanket commercial or redistribution license is asserted.

Files with SHA256: 202a1c27917990e039c8421a183339629acd985796a70ef26e9e443f65b916ea