wesjos/Qwen-Image2.1-Genshin-2D-Lora

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

---

🎨 Gallery — text-to-image (t2i)

All six samples are native generations (no upscaling, no inpainting) at genshintcg_style strength 0.8 — 1–3 at 832×1248, 4–6 at 944×1104.

---

🎨 Gallery — image editing (edit / img2img)

prompt: Reapply this image using the genshintcg_style.

1 · Photoreal → 2D character art (strongest example)

2 · Anime illustration → vivid cel-shaded version

3 · Anime illustration → high-contrast version

---

📋 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
0.8	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	5.0
Sampler	euler
Scheduler	simple
Shift (ModelSamplingAuraFlow)	3.0
Size	832 × 1248 (t2i gallery) / follows input for edit

Negative prompt

deformed, wrong hands, mutated, fused fingers, extra hand, extra feet, ugly underwear

---

⚠️ Limitations

1. Trained mainly on single-character shots. The gallery is portraits and half-body; multi-character scenes and complex full-body compositions generalize less reliably.
2. 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.
3. Composition drift rises with edit denoise. Above roughly 0.85 denoise the output starts re-arranging the subject rather than restyling it.
4. Above ~1.0 strength the style over-compresses eyes, hands and layered clothing; drop the strength instead of pushing it.
5. 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.
6. 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.