ComfyPilotComfyPilot
模型库 / LoRA / MOSS-Image Edit CharRef (beta)
MOSS-Image Edit CharRef (beta)

MOSS-Image Edit CharRef (beta)

3011337

LoRA 384 次下载 28 喜欢 更新于 2026-11-17

底模:MOSS Image

animeconceptmoss-imageimage editingconsistent characterreference

下载模型 在源站打开

文件清单 线路:镜像(hf-mirror.com) · 切换

下载后放进对应的 ComfyUI 目录。models/ 指 ComfyUI 的模型根目录。

权重 moss_image_edit_charref_v2.safetensors 284 MB 下载
权重 moss_image_edit_charref_v2_node_and_workflows.zip 617 KB 下载

说明

MOSS-Image Edit CharRef (beta)

作者:bag3l

An edit LoRA for MOSS-Image SFT: Picture 1 is an anime character, the prompt describes a new scene and pose, and the model draws the same character there. Hair, face, outfit and accessories come from the picture. Pose, framing and background come from the prompt. It is not tied to any particular character: any anime character can be the reference. Bagel the fox and Kujou, in the gallery, are only examples. It stacks with style LoRAs. This LoRA carries the character and has no strong look of its own, so the look can come from your style LoRA at 0.6 to 0.8. The two sheets below show it. ⚠️ Still an early version. On 24 test characters that were not in training, an automatic same-character check passes about half of the plain renders and 61% with reference guidance. Small details (skin tone, eye details, small accessories) are often lost, and results vary between seeds. Anime characters only. v2 replaces the first experiment (v1). The two versions need different models, nodes and prompts. Everything below is for v2; the v1 notes are at the bottom. With a style LoRA Each sheet has the two references on the left and six scenes across. The first row is this LoRA alone. The rows below add the style LoRA bag3l_style at 0.6 and at 0.8, on two seeds. Every render is shown, nothing was removed.

  • The character holds. The same-character check passed 48 of the 48 styled pictures, and 11 of the 12 without a style. That is two characters with one reference each, so read it as a first check and not a benchmark.
  • 0.6 to 0.8 is a good strength. At 0.8 the look is stronger and the character drifts a little more.
  • Check the hands, and watch for extra props. With the style on, a few full-body pictures gained a handbag that is in neither the reference nor the prompt. Try another seed.

These sheets were rendered with MOSS-Image SFT Turbo at 8 steps. The style LoRA in them, bag3l_style, is not publicly available. It is the only style LoRA tested here; others may behave differently. What you need (ComfyUI)

  • A current ComfyUI that has the Text Encode MOSS-Image Edit node
  • Custom node moss_image_charref: in the Archive file on this page (zip), or on GitHub. Copy the folder into ComfyUI/custom_nodes/ and restart. It needs no extra packages. The ostris node that v1 used is not needed.
  • From huggingface.co/Comfy-Org/MOSS-Image:
  • moss_image_sft_bf16.safetensors → models/diffusion_models (SFT, not Base)
  • qwen3.5_9b_bf16.safetensors → models/text_encoders
  • flux2-vae.safetensors → models/vae
  • This LoRA → models/loras, named moss_image_edit_charref_v2.safetensors
  • A workflow, from the same zip:
  • moss_charref_reference_guidance.json: the closest likeness. It runs one extra pass without the reference, so each image takes about 1.4 times as long.
  • moss_charref.json: plain sampling, faster.
  • moss_charref_turbo_style.json: MOSS-Image SFT Turbo at 8 steps, with a slot for your style LoRA. It needs moss_image_sft_turbo_int8_convrot.safetensors from toxicdog/MOSS-Image-SFT-Turbo-INT8-ConvRot and qwen3.5_9b_int8_convrot.safetensors from Comfy-Org.

The custom node is required. ComfyUI normally places a reference picture on the same positions as the output, and this LoRA was trained with the reference placed below the output. Without the node the LoRA does not work as tested. How to use

  • Reference: one character on a plain white background, in a neutral pose with the arms down. A full-body picture keeps the whole outfit, the tail and the shoes. A picture cropped from head to thighs keeps the face, eye colour and ears better. A clean, flat-shaded picture gives the cleanest results. The workflow fits it into a 448 px white square for you.
  • Prompt: comma-separated tags for the scene, pose, framing and expression, starting with 1girl, solo or 1boy, solo. There is no trigger sentence. Unlike v1, leave the character's appearance tags out: the looks come from the picture. 1girl, solo, sitting, cafe, table, holding cup, coffee, window, sunlight, smile, looking at viewer, upper body
  • Settings on MOSS-Image SFT (already set in the first two workflows): 50 steps, CFG 3, euler / sgm_uniform, skip layer 22, edit format ai-toolkit, LoRA 1.0, output about 0.6 MP, for example 672 x 896. Reference guidance 2.5 in the guided workflow.
  • Settings on MOSS-Image SFT Turbo: 8 steps, CFG 1, euler / simple, the same size. Keep reference guidance off or at 1.5 at most; at 2.5 it copies the reference's pose.
  • Style LoRA: load it right after this one, before the Reference layout node, at 0.6 to 0.8. Put its trigger word, if it has one, at the start of the prompt.
  • Leave the negative prompt empty.

Tip: when the character comes out wrong, try another seed first, then the reference guidance workflow. Results Measured on MOSS-Image SFT at 50 steps, without a style LoRA: 24 anime characters that were not in training, each with a prompt for a new pose and scene, on three seeds, 72 images per line.

  • Plain: same character kept in 50%, pose and scene follow the prompt in 85%
  • Reference guidance 2.5: same character kept in 61%, pose and scene follow the prompt in 78%

"Same character kept" is an automatic same-character check (CCIP) between the output and the reference. It is strict: an image can look right to a person and still fail it, and the reverse. MOSS-Image SFT Turbo and the style LoRA stack were checked by eye and with the small check above, not with this test. On an RTX 5090, MOSS-Image SFT took about 19 s an image plain and 26 s with reference guidance. Limitations Anime characters only, not for real people. One character and one reference.

  • Details get lost. Skin tone, eye details and small accessories are the first to go. The last two gallery images of v2 show it: the boy's tan skin came out pale and his golden eyes turned into tinted glasses, and the maid's eyepatch became hair over the eye or a yellow lens.
  • The reference's pose can leak. A reference with raised hands can give a picture with four hands: the reference's and the prompt's. A negative prompt did not fix it in our test. Use a neutral reference.
  • It copies the reference's rendering. A painterly or effect-heavy reference pulls its brushwork, colour cast and effects into the output, which can come out muddy. A reference sheet with two figures can give two figures.
  • Things can be added, such as a forehead gem on the elf in one gallery scene.
  • Not tested: MOSS-Image Base, sizes above about 0.6 MP, more than one reference picture, photos, and other LoRA strengths or samplers.
  • The numbers come from a single training run.

The v2 gallery shows the two style sheets and eight original characters, all that were rendered for that set: six that worked and the two misses, each with its reference and three scenes at seed 1. Older version: v1 (early experiment) The v1 file on this page is for MOSS-Image Base and works differently. It is kept for reference; v2 is the better one.

  • It needs the custom node ComfyUI-MOSS-Image-Ostris-Edit by ostris (github.com/ostris/ComfyUI-MOSS-Image-Ostris-Edit), moss_image_base_int8_convrot.safetensors and its own workflow, the Archive file of the v1 version.
  • Its prompt has three parts: Same character and art style as Picture 1., the scene tags, and the character's own appearance tags.
  • Settings: 30 steps, CFG 3, euler / simple, edit patch kv_cache on, skip layer 22, output about 1 MP, LoRA 1.0.
  • It was a first experiment: 2,000 steps on about 830 image pairs from 16 characters.

Credits OpenMOSS (MOSS-Image, Apache-2.0), ostris (AI Toolkit edit training), Comfy-Org (ComfyUI and the model files), toxicdog (the SFT Turbo int8 model). Outputs are AI-generated, so please label them that way. Node, workflows, full details and license (MIT): github.com/bag3labo/moss-image-edit-charref-v2

示例图