Genshin TCG 2D Style LoRA — Qwen-Image-2.1
Style LoRA for
Qwen-Image-2.1. Works in both text-to-image and image-editing (img2img / image edit) mode. trained on comfyui's qwen_image_2.1_int8_convrot.safetensors, pls use this as based model Trigger word:genshintcg_style
🖼️ Training Data — style references
The style was learned from Genshin Impact 2D character art. These are sample images from the training set — the rendering the LoRA reproduces: flat cel shading, painted fabric and metal detail, coloured linework, soft gradient backdrops.
Ganyu cool blue / teal palette lotus-and-lake motif |
Nahida green / gold palette dune-and-leaf motif |
Xiangling warm orange / cream palette flame-and-cloud motif |
🎨 Gallery — text-to-image (t2i)
Nine native generations — no upscaling, no inpainting. Same trigger word, one prompt per picture; the only thing that changes between rows is framing and light. Every cell lists subject · palette · native resolution (click opens full size).
① Full-body, plain and sunlit backdrops — 832 × 1248
Seated, white backdrop silver-blue curls · white tee · denim shorts 832×1248 |
Sunlit outdoor, waving blonde braid · black two-piece · pale sky 832×1248 |
Holding a cat black ponytail · headphones · denim shorts 832×1248 |
② Square character shots — 1024 × 1024
Sport scene blue-and-gold costume · basketball · bright hall 1024×1024 |
Nahida, forest light green sleeveless dress · sun through leaves 1024×1024 · seed 2085735429 |
Close-up, raised hand aqua braid · white tank with green trim 1024×1024 · seed 2087929872 |
③ Warm interiors — 1024 × 1024, 1216 × 1216 and one 1600 × 1024 wide
Mai Shiranui, shoji light red-white wrap · gold sash 1024×1024 · seed 2088672379 |
Asuna, classroom window blonde braid · white blouse 1216×1216 · seed 2089509578 |
Low-key interior silver-blue curls · black tank 1600×1024 |
Recipe: t2i_01–03, 07, 08 at genshintcg_style strength 0.8; t2i_09–12 at 0.75, CFG 4.0, euler / simple, 30 steps, ModelSamplingAuraFlow shift 3.0 — those four ship their full ComfyUI graph inside the PNG (prompt text chunk), so every number quoted here can be read back from the file itself. Style checkpoint for the newer set: genshintcg_style_st4000.safetensors.
🎨 Gallery — image editing (edit / img2img)
prompt: Reapply this image using the genshintcg_style.
· Asuna
📋 Model Details
| Property | Value |
|---|---|
| Base model | Qwen-Image-2.1 (qwen_image_2) |
| Type | Style LoRA — Genshin-style 2D character art |
| Trigger word | genshintcg_style |
| Modes | text-to-image and image editing (img2img) |
| Rank | 32 |
| Tensors | 384 (192 pairs of lora_A / lora_B) |
| Target modules | Per layer: attn.to_q · to_k · to_v · to_out.0 · img_mlp.gate_up · img_mlp.out |
| Blocks covered | All 32 layers transformer_blocks.0–31 |
| Dtype | F16 |
| Alpha | Not written → loaders treat as alpha = rank = 32 (scale 1.0) |
| Training framework | ai-toolkit v0.13.21 · flowmatch |
| Key format | ComfyUI-style diffusion_model.* prefixes (ai-toolkit save format) |
| Size per file | 152 MiB each |
📦 Checkpoint Guide
| File | Step / Epoch | Size | Recommendation |
|---|---|---|---|
genshintcg_style_st3000.safetensors |
3000 / 27 | 152 MiB | ✅ Safe, most restrained colouring |
genshintcg_style_st3500.safetensors |
3500 / 31 | 152 MiB | ⭐ Recommended default — best balance of linework and palette |
genshintcg_style_st4000.safetensors |
4000 / 36 | 152 MiB | ✅ Strongest saturation, most stylised |
The three checkpoints ship here; no weight is excluded from the guide.
🚀 Usage
Text-to-image
genshintcg_style, a girl with long purple hair sitting on the floor, white t-shirt, blue shorts, plain white background
Image editing (img2img)
The prompt for edit mode is the edit instruction, not a scene description — the input image already carries the content.
genshintcg_style, 2d character art style
Feeding a full scene description in edit mode competes with the input image and pushes the result toward a new composition. Keep the prompt short.
LoRA strength
| Scale | Effect |
|---|---|
| 0.6 – 0.7 | Mild stylisation; retains more of the input's original rendering |
| 0.75 – 0.8 | Sweet spot used for the gallery above — linework and palette clearly present without flattening detail |
| 0.9 – 1.0 | Heavier stylisation; small facial features and fabric patterns start to compress |
ComfyUI
Base model files (diffusion_models + text_encoder + vae):
| Slot | File |
|---|---|
| diffusion model | qwen_image_2.1_int8_convrot.safetensors |
| text encoder | qwen3vl_8b_int8_convrot.safetensors |
| VAE | qwen_image_2.1_vae_bf16.safetensors |
KSampler parameters (measured from this repo's example images):
| Parameter | Value |
|---|---|
| Steps | 30 |
| CFG | 4.0 – 5.0 (4.0 on t2i_09–11, read from the graph embedded in those PNGs) |
| Sampler | euler |
| Scheduler | simple |
| Shift (ModelSamplingAuraFlow) | 3.0 |
| Size | 832 × 1248 / 1024 × 1024 / 1216 × 1216 / 1600 × 1024 (t2i gallery) · follows input for edit |
Negative prompt
deformed, wrong hands, mutated, fused fingers, extra hand, extra feet, ugly underwear
⚠️ Limitations
- Trained mainly on single-character shots. The gallery is portraits and half-body; multi-character scenes and complex full-body compositions generalize less reliably.
- Photoreal inputs in edit mode lose photographic texture by design — skin pores, lens effects and fabric weave are replaced by painted shading. Subject, pose and framing survive; photorealism does not.
- Composition drift rises with edit denoise. Above roughly 0.85 denoise the output starts re-arranging the subject rather than restyling it.
- Above ~1.0 strength the style over-compresses eyes, hands and layered clothing; drop the strength instead of pushing it.
- Generated content is not official Genshin lore. Costume and character details may not match the source material — defer to official assets for anything canon-critical.
- 8-bit lines need real resolution. Inputs below ~512 px on the short side produce soft, muddy output.
📄 License
Apache-2.0, matching the Qwen-Image-2.1 base model.