🎬 CineWatch puts Overwatch heroes into a cinematic CGI look: film-style lighting, real depth of field, moody shadows and detailed backgrounds, like frames from a cinematic trailer.

📚 An Anima checkpoint trained on 432 curated cinematic renders of 25 Overwatch heroes. Every image is captioned with booru tags plus a full sentence, so it understands both prompt styles.

✨ What it does well

• 🦸 Heroes on-model by name, no trigger word: tracer \(overwatch\), overwatch gives Tracer with her outfit and gear.

• 💡 Lighting that follows your prompt: backlight and silhouettes, neon, firelight, storms, low-key dark scenes.

• 🎯 Keeps Anima's prompt following: colours, left/right placement, counts and short text still work.

• 🌐 Non-Overwatch prompts stay neutral. Other characters come out in the same 3D CGI look, and adding anime screencap, anime coloring, flat color, cel shading, 2d gives clean 2D anime 🎨

🦸‍♀️ Heroes in the dataset

Ana, Anran, Ashe, Brigitte, D.Mon, D.Va (and MEKA), Domina, Freja, Illari, Juno, Junker Queen, Kiriko, Mei, Mercy, Moira, Pharah, Shion, Sierra, Sojourn, Sombra, Symmetra, Tracer, Vendetta, Widowmaker, Zarya.

The dataset is mostly women; male heroes come from Anima's own knowledge.

⚙️ How to use (ComfyUI)

• 🧠 Load Diffusion Model: CineWatch_AnimaV1.safetensors (diffusion model only, bf16)

• 📝 Text encoder: qwen_3_06b_base.safetensors (Load CLIP, type stable_diffusion)

• 🖼️ VAE: qwen_image_vae.safetensors

• Same text encoder and VAE as the Anima base model.

• 🎛️ ModelSamplingAuraFlow shift 3 · 30 steps · CFG 4 · Euler · Simple

• 📐 Around 1.3 MP: the gallery uses 864x1536 and 1536x864

✍️ Prompt format

masterpiece, best quality, score_7, safe, 1girl, solo, <hero> \(overwatch\), overwatch, <tags>. <One or two sentences describing the moment, the camera angle and the light.>

💬 Escape the brackets in hero tags in ComfyUI: kiriko \(overwatch\).

🎥 Tip: naming the camera angle and the light source in the sentence ("worm's-eye view", "lit only by the forge", "backlit by floodlights") makes the biggest difference.

🚫 Negative prompt

worst quality, low quality, score_1, score_2, score_3, blurry, jpeg artifacts, sepia, bad anatomy, bad hands, extra fingers, missing fingers, extra digit, fewer digits, extra limbs, deformed, watermark, text, signature, artist name, logo

⚠️ Known limits

• "cowboy shot" often still gives a full body.

• Lettering on clothes (jerseys, logos) comes out garbled.

• Domina tends to get sneakers instead of her silver boots.

• Two-person action (who holds the ball, who blocks) can take a few seeds.

📸 Every gallery image has its full prompt and settings attached.