═══════════════════════════════════════════════════════════════════════════════ ANIME2REAL2511-PAULHE-V2.0 ═══════════════════════════════════════════════════════════════════════════════ ─────────────────────────────────────────────────────────────────────────────── IDENTITY ─────────────────────────────────────────────────────────────────────────────── Base model Qwen-Image Generation mode not declared by the author Checkpoint not determined — verify before use Category Physical effect / transformation Author paulhe Source https://civitai.com/models/2657054 Provenance CivitAI · model 2657054 · version 2983568 ─────────────────────────────────────────────────────────────────────────────── WEIGHTS ─────────────────────────────────────────────────────────────────────────────── File anime2real-paulhe-v2.0.safetensors Version v2.0 Size 281.5 MB Format SafeTensor / unknown precision SHA256 7dc1c37d9af146d83fd8245779806efe026ab5c22ba661faea1c3f459d5a1db3 Published 2026-05-28 ─────────────────────────────────────────────────────────────────────────────── TRIGGER WORDS ─────────────────────────────────────────────────────────────────────────────── No trigger word. The LoRA applies as soon as it is loaded — strength is the only way to dial the effect up or down. ─────────────────────────────────────────────────────────────────────────────── RECOMMENDED SETTINGS ─────────────────────────────────────────────────────────────────────────────── The author did not publish specific settings. No author settings were published. Community baseline: cfg 4.0 without an accelerator - `true_cfg_scale=4.0` in Qwen's own diffusers snippet for Qwen-Image-2512, and cfg 4.0 in the Comfy-Org templates image_qwen_Image_2512, image_qwen_Image_2512_controlnet and template_qwen_Image_2512_360_lora. cfg 1.0 whenever a Lightning/Turbo LoRA is loaded (ModelTC README CLI examples `--steps 8 --cfg 1.0` / `--steps 4 --cfg 1.0`; the base image_qwen_image, 2-step Turbo and union-control templates all sample at cfg 1.0). Editing: the Edit-2511 template uses cfg 3.0 with a CFGNorm node; the Edit-2509 template uses cfg 1.0 with a 4-step Lightning LoRA. Illustration-LoRA template: cfg 3.5. Layered template: cfg 2.5. CivitAI's sdcpp Qwen worker defaults to cfgScale 2.5 (developer.civitai.com notes "Qwen wants ~2.5")., lora_strength 1.0. Every official Comfy-Org Qwen template loads its LoRAs at strength 1.0 - verified in image_qwen_image (Lightning 8-step), image_qwen_Image_2512, image_qwen_Image_2512_controlnet, image_qwen_image_2512_with_2steps_lora, image_qwen_image_edit_2509, image_qwen_image_edit_2511, image_qwen_image_union_control_lora (two LoRAs, both 1.0), template_qwen_Image_2512_360_lora (two LoRAs, both 1.0) and template_qwen_image_illustration_lora. Community practice for stacking concept LoRAs is to keep the summed strength around 0.9-1.2 (CivitAI article 18798 "Qwen-Image NSFW LoRA Notes": "The sum should be in the range of 0.9 to 1.2")., resolution Official aspect-ratio table, identical in the Qwen-Image and Qwen-Image-2512 model cards: 1:1 1328x1328, 16:9 1664x928, 9:16 928x1664, 4:3 1472x1104, 3:4 1104x1472, 3:2 1584x1056, 2:3 1056x1584. 1328x1328 is the EmptySD3LatentImage default in the base, 2512, 2512-ControlNet and 2-step Turbo templates. Exceptions in official templates: the 360-panorama template uses 2048x1024, the illustration-LoRA template 1024x1024, the Layered template 640x640. CivitAI's sdcpp Qwen worker defaults to 1024x1024., sampler euler + simple, denoise 1.0 - verified as the sampler/scheduler/denoise triple in every Comfy-Org Qwen template inspected (base, 2512, 2512 ControlNet, 2-step Turbo, Edit-2509, Edit-2511, union control, 360 panorama, illustration LoRA, Layered). All of them also insert a ModelSamplingAuraFlow node before the sampler, because ComfyUI's built-in shift for the qwen_image model type is only 1.15 (comfy/supported_models.py). The shift value is NOT uniform: 3.1 in image_qwen_image, image_qwen_Image_2512, image_qwen_Image_2512_controlnet, template_qwen_Image_2512_360_lora, image_qwen_image_union_control_lora and image_qwen_image_edit_2511; 3.0 in image_qwen_image_edit_2509 and image_qwen_image_2512_with_2steps_lora; 7.0 in template_qwen_image_illustration_lora; 1.0 in image_qwen_image_layered. Treat 3.1 as the T2I default, not a universal constant., steps 50 for Qwen- Image-2512 without an accelerator (official diffusers `num_inference_steps=50` and the KSampler default in image_qwen_Image_2512, image_qwen_Image_2512_controlnet and the 360 template). Editing: 40 steps in the Edit-2511 template; 4 steps in the Edit-2509 template (which ships with a 4-step Lightning LoRA). Illustration-LoRA template: 45. Layered template: 20. With accelerators the step count is exactly what the LoRA filename says: 4 or 8 for Lightning, 2 for the Wuli 2-step Turbo, always at cfg 1.0. Community full- distill checkpoint: 15 steps at cfg 1.0 per ComfyUI docs, acceptable at 10.. ─────────────────────────────────────────────────────────────────────────────── ADOPTION & POPULARITY ─────────────────────────────────────────────────────────────────────────────── Downloads 538 Rating 100% positive (43/43) Comments 2 Published 100 days ago Download rate 5.4/day since release Adoption rank #191 of 200 in this collection (Niche) Momentum rank #62 of 200 by download rate (Steady) Adoption tier is the percentile of total downloads within this collection; momentum is the percentile of downloads-per-day since release. Momentum is ranked rather than measured against a fixed rate, because this base model is itself new and every LoRA looks fast on an absolute scale. Both numbers are CivitAI's, read on the scrape date at the foot of this file. ─────────────────────────────────────────────────────────────────────────────── CHAINING / STACKING ─────────────────────────────────────────────────────────────────────────────── Stacks with others not stated The author gave no chaining guidance. The rules below are the general Qwen-Image ones, not anything specific to this LoRA. General Qwen-Image rules that apply here (see REFERENCE.txt): • Stack by wiring LoraLoaderModelOnly nodes in series on the MODEL line. This is the officially shipped pattern: template_qwen_Image_2512_360_lora chains Qwen- Image-2512-Lightning-4steps-V1.0-fp32 at 1.0 into qwen-360-diffusion-2512-int8-bf16-v2 at 1.0. • One accelerator LoRA maximum. Every official template loads exactly one Lightning/Turbo LoRA, and ComfyUI's Qwen-Image docs record that the fully-distilled checkpoint and the lightx2v 8-step LoRA are not usable together. • Match the accelerator to the checkpoint family: Qwen- Image-2512-Lightning on 2512, Qwen-Image-Lightning on the August base, the Edit-2509 subfolder LoRAs on Edit-2509, Qwen-Image- Edit-2511-Lightning on Edit-2511. lightx2v distilled and shipped these separately per family. • When any accelerator LoRA is in the chain, move the sampler to that LoRA's step count and cfg 1.0. Leaving cfg at 4.0 with a distilled LoRA is the most common bad-output cause. • Order used by the official templates: accelerator LoRA adjacent to the base loader, concept/style LoRA after it, then ModelSamplingAuraFlow, then the sampler (the 360-panorama template's exact order). • Keep the summed strength of concept/style LoRAs roughly in the 0.9-1.2 band (CivitAI article 18798). The accelerator LoRA sits outside that budget at its own 1.0. • Control LoRAs (qwen_image_union_diffsynth_lora.safetensors) chain like ordinary LoRAs and combine with an accelerator: the official 'Qwen- Image Union Control' template stacks the control LoRA with Qwen-Image- Lightning-4steps-V1.0, both at 1.0, sampling at 4 steps / cfg 1.0 / shift 3.1. • Be conservative stacking older Edit LoRAs on Edit-2511: the Qwen README says community LoRA effects were merged into the 2511 base, so those LoRAs can double-apply. • The author never states which weight family this targets. Mismatch is silent degradation, essentially never a hard error. Every Qwen-Image weight family resolves to the same ComfyUI model class (comfy.model_base.QwenImage, unet_config image_model "qwen_image", latent_format Wan21, built-in sampling shift 1.15 - comfy/supported_models.py ─────────────────────────────────────────────────────────────────────────────── PROS ─────────────────────────────────────────────────────────────────────────────── + Compact at 281 MB — cheap to stack with other LoRAs. [source: file size] ─────────────────────────────────────────────────────────────────────────────── CONS & CAVEATS ─────────────────────────────────────────────────────────────────────────────── - No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata] - Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description] - Modest adoption (538 downloads) — less community feedback to rely on. [source: source stats] ─────────────────────────────────────────────────────────────────────────────── COMPATIBILITY WARNING ─────────────────────────────────────────────────────────────────────────────── Mismatch is silent degradation, essentially never a hard error. Every Qwen- Image weight family resolves to the same ComfyUI model class (comfy.model_base.QwenImage, unet_config image_model "qwen_image", latent_format Wan21, built-in sampling shift 1.15 - comfy/supported_models.py), and the bf16/fp8 file sizes are identical across Qwen-Image, Qwen-Image-2512, Qwen-Image-Edit, Edit-2509 and Edit-2511. So a LoRA trained on any of them will key-match and load against any other. Where keys genuinely do not match, ComfyUI only logs `lora key not loaded: ` (comfy/lora.py:93) and continues; a shape mismatch logs `WARNING SHAPE MISMATCH WEIGHT NOT MERGED` (comfy/lora.py:498) and skips that tensor. Both are console warnings; the render still completes. Verified consequence: lightx2v re-distilled and re-released Lightning weights separately for Qwen-Image, Qwen-Image-Edit, Edit-2509, Edit-2511 and Qwen-Image-2512, so accelerators are family- specific in practice. The one documented case worse than mild degradation is precision-level, not family-level: bf16-trained LoRAs on the naive qwen_image_fp8_e4m3fn.safetensors base produce grid artifacts, stated outright in the ModelTC/Qwen-Image-Lightning README. Whether an Edit LoRA on the T2I base merely no-ops (no ref_latents conditioning is supplied) or a T2I LoRA on an Edit checkpoint transfers as a style push is inference from the architecture, not a sourced statement - see unverified. ─────────────────────────────────────────────────────────────────────────────── AUTHOR'S DESCRIPTION ─────────────────────────────────────────────────────────────────────────────── Yes! I trained another LoRA again. Although the previous LoRA already restored expressions quite well, this one has even better expression fidelity than the last LoRA. As for consistency, I think it ranks very highly among all real-life conversion LoRAs. Feel free to test it out. I tested it on multiple ahegao images as well as some subtle micro-expressions. In most cases, the expressions can be restored very well with the LoRA alone. For a small number of images, you can try optimizing the prompt to achieve a better restoration. Here, I used LLaMA to reverse-engineer the prompts and fed them into Qwen 2511. I simply added a prompt reverse-engineering module to the official example template. I am only providing the LoRA model here. If you need the workflow, you can also ask me. I may release the workflow later as well. If you want better skin quality, you can connect it with Z-Image Turbo for image refinement. I should upload the related workflow to RunningHub, so you can try it there. This LoRA was also trained on an RTX 5090, using 69 images for 7,500 training steps. Enjoy the power of the 5090! trigger words: transform into a real-life version. Strictly preserve the facial expression and pupil color, and strictly keep the hairstyle unchanged. with facial features like those of a real person. Make the skin more realistic, with fine peach fuzz. Keep the character’s proportions in the image unchanged. Strictly preserve the character’s expression and subtle detailed movements. If you feel that the result generated with the official workflow alone is not accurate enough, you can try optimizing the prompt after my original prompt using the same sentence structure. However, it is best not to delete my original prompt. This is especially useful for images with subtle micro- expressions. If you think the result is good, please follow and give it a like! 是的!我又练了一炉lora,虽然上一次lora表情已经比较还原,但是这次的表情还原性比上一个lora还要好一点,一致性我认为在所有的转真人lora里面 也算是排很前的了。你们不妨测试一下,我测试了多个阿黑颜照片,还有一些微表情,绝大部分单靠lora就能很好还原,有一小部分可以尝试优化提示词来还原。这里 我使用的是用llama反推提示词喂给qwen2511。而我只是在官方的示例模板里面加入了一组反推模块而已。这里我只提供lora模型,如果你们需要工作流 ,也可以向我询问,我可能后续也会发布工作流。 如果想要皮肤质量更好的话可以接zimageturbo进行洗图,相关工作流我应该会上传到runninghub上面,你们可以试试。 这次同样是使用rtx5090进行训练,使用69张图片进行7500步的训练,尽情享受5090的威力吧! 提示词:transform into a real-life version. Strictly preserve the facial expression and pupil color, and strictly keep the hairstyle unchanged. with facial features like those of a real person. Make the skin more realistic, with fine peach fuzz. Keep the character’s proportions in the image unchanged. Strictly preserve the character’s expression and subtle detailed movements. 如果觉得单纯用官方工作流生成的有一点点不还原,你可以尝试在我这个原来提示词后面用同样的句式优化提示词。但是注意最好不要把我原来的提示词删掉。这在一些微 表情的图中尤其管用。 如果觉得效果不错就点一下关注和点赞吧! ─────────────────────────────────────────────────────────────────────────────── VERSION NOTES ─────────────────────────────────────────────────────────────────────────────── Enhanced training for consistency, resulting in more faithful restoration of the character’s expressions. ─────────────────────────────────────────────────────────────────────────────── EXAMPLE MEDIA ─────────────────────────────────────────────────────────────────────────────── 6 preview(s) in ./example_images/ — community results for this LoRA. Videos keep a matching _poster.jpg still frame. No prompts are recorded below: CivitAI's public API flags these posts as having generation metadata but does not return it, so the prompts cannot be scraped. Open the model page to read them. 01_image.jpeg (image, 2162x1448) 02_image.jpeg (image, 1306x360) 03_image.jpeg (image, 2162x1448) 04_image.jpeg (image, 1014x279) 05_image.jpeg (image, 2048x1024) 06_image.jpeg (image, 2048x1024) ─────────────────────────────────────────────────────────────────────────────── HOW TO USE (COMFYUI) ─────────────────────────────────────────────────────────────────────────────── 1. Update ComfyUI first - Qwen-Image support is native, but each new family member (Edit-2509, Edit-2511, 2512) needed a fresh release, so an old build will not detect the checkpoint. 2. Put the base checkpoint in `ComfyUI/models/diffusion_models/` - e.g. `qwen_image_2512_fp8_e4m3fn.safetensors` (20.43 GB, the file the official 2512 template names) or `qwen_image_2512_bf16.safetensors` (40.86 GB), from huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/split_files/diffusion_models. Edit checkpoints come from Comfy-Org/Qwen-Image-Edit_ComfyUI. Load with `Load Diffusion Model` (UNETLoader), not a checkpoint loader - these are unet-only files. 3. Put `qwen_2.5_vl_7b_fp8_scaled.safetensors` (9.38 GB) in `ComfyUI/models/text_encoders/` and load with `Load CLIP` (CLIPLoader), type `qwen_image`. Every official template uses exactly this pair. 4. Put `qwen_image_vae.safetensors` in `ComfyUI/models/vae/` and load with `Load VAE`. 5. Put LoRA `.safetensors` in `ComfyUI/models/loras/`. Subfolders show up as `subfolder/name.safetensors` in the dropdown - useful for separating t2i / edit / accelerator LoRAs. 6. Load with `LoraLoaderModelOnly`, patching only the MODEL line between the UNETLoader and ModelSamplingAuraFlow. Every official Comfy-Org Qwen template uses LoraLoaderModelOnly; the two-output `LoraLoader` also works but its CLIP branch is a no-op for DiT-only LoRAs. 7. Insert `ModelSamplingAuraFlow` after the LoRA chain. Use 3.1 for T2I and Edit-2511, 3.0 for the 2-step Turbo and Edit-2509 templates. Without this node ComfyUI uses the model's built-in shift of 1.15 and results differ from every published example. 8. Set KSampler to euler / simple / denoise 1.0, then steps+cfg per the accelerator (50/4.0 plain, 8/1.0 or 4/1.0 with Lightning, 2/1.0 with the 2-step Turbo). Latent from `EmptySD3LatentImage` at an official resolution. 9. Low-VRAM alternative: the `ComfyUI-GGUF` custom node with `Unet Loader (GGUF)` and a Qwen-Image/2512 GGUF (unsloth/Qwen-Image-2512-GGUF). LoRA loading still goes through LoraLoaderModelOnly, placed after the GGUF loader. 10. For Edit workflows, swap CLIPTextEncode for `TextEncodeQwenImageEdit` (single image) or `TextEncodeQwenImageEditPlus` (2509/2511) and add a `CFGNorm` node as the Edit-2511 template does; the LoRA chain sits in the same place. ─────────────────────────────────────────────────────────────────────────────── Scraped by lorakit on 2026-09-05 03:55 UTC. Stats and description are the author's; pros/cons are derived from the evidence tagged beside each line. ───────────────────────────────────────────────────────────────────────────────