Epic Cinematic Anime
Style LoRA for CircleStone Anima that turns any prompt into a climax shot.
Sharp, jagged edges everywhere: angular shapes, spiky ink lines and hard diagonals over dry painterly brushwork, back-lit dramatic lighting, fire against deep shadow, and a camera that never stands still.
Most style LoRAs change how the picture is rendered and keep the base model's composition. This one also re-stages the scene: lower angles, stronger foreshortening, the monster leaning over the hero, the hammer swung straight at the camera. Same prompt, same seed, a different film.
Monsters are its specialty. Dragons, kaiju, demons and forest beasts come out huge, toothy and scary, with glowing eyes and inked scales. Put a tiny figure next to them for scale.
Works on any character the base model knows (Frieren, Makima, Gojo, Miku, Asuka, Mikasa…) and on calm scenes too: a cat on a windowsill or a girl in a wheat field get the same light and brushwork.
Shines in wide formats. The source art is mostly cinematic 21:9, so 1536×640 and 1344×768 give the strongest results. The gallery is vertical, so open the wide images to see them uncropped.
Inspired by the art of @sonic1nstinct. Not an official model by the artist; please support the original artist.
────────────────────────────────────────────────────────────
Anima
Base model: Anima (anima-base-v1.0)
Trigger: @snkbrush (at the very start of the prompt, followed by a comma, then regular Danbooru tags)
No style tags needed.
LoRA strength: 1.0
Example prompt
@snkbrush, safe, masterpiece, best quality, score_7, 1girl, solo, emilia (re:zero), white hair, long hair, purple eyes, pointy ears, white dress, ice, snow, casting spell, blue theme
@snkbrush, safe, masterpiece, best quality, score_7, no humans, kaiju, giant monster, rising from sea, glowing eyes, roaring, storm, lightning, waves, ships, night
Recommended settings
Sampler: DPM++ 2M SDE, scheduler simple, 25 steps, CFG 5
Size: 1536×640 (21:9) or 1344×768 (16:9) for cinematic shots, 832×1248 (2:3) for portraits
Hires fix: 1.5×, lanczos, 20 steps, denoise 0.4. This is where the ink scratches and brush texture really come out
Negative prompt:
worst quality, low quality, score_1, score_2, score_3, jpeg artifacts, logo, watermark, signature, twitter username, bad anatomy, bad hands, missing finger, extra digits, fewer digits, nsfw, nude, explicit, cleavage
Tips
Ask for the shot: from below, dynamic pose, wide shot, small figure, looking up, light rays, sparks, smoke, rain. The LoRA pushes them further than the base model does.
Monsters: dragon, giant monster, kaiju, demon, glowing eyes, sharp teeth plus 1girl, small figure, standing for scale. no humans for a creature portrait.
Break the warm palette: the LoRA leans to sunset orange and sepia. night, rain, blue theme, ice, neon lights, moonlight give cold blue and teal frames.
────────────────────────────────────────────────────────────
Known issues
It sometimes trades the face for the angle: characters turn their back to the camera. Add looking at viewer or reroll.
Busy battles can break weapons and chains (a sword turning into two, a flail chain snapping). Simpler action reads better.
Signs and neon text come out as gibberish.
Trained on Anima Base v1.0 only. Behaviour on Anima Turbo and third-party merges is unknown.
Showcase images use hires fix 0.4. Full prompts are in the image metadata. Before/after pairs use the same prompt and seed without the LoRA.
────────────────────────────────────────────────────────────
Training
50 digital originals (mostly 4K, 21:9), 2 duplicates dropped; letterbox bars trimmed; every picture kept whole (≤1600 px) for the wide composition, and pictures wider than 2:1 also cut into overlapping 3:2 tiles so the brushwork is seen near 1:1 → 117 images
Short captions: trigger + rating + character names (WD tagger ≥0.85 minus misreads, plus names checked by hand) + people count, framing, hair colour, a few items + subject tags (dragon, duel monster, monster, giant, blood, fire, battle) so content stays out of the trigger; nothing about style, colours or lighting, no quality tags
Anima LoRA, rank 32 / alpha 32, incl. AdaLN modulation; DiT trained, text encoder and LLM adapter frozen · AdamW8bit, LR 1e-4, constant with 5% warmup · 10 epochs + 2 more continued from epoch 10 = epoch 12 (≈1400 steps), batch 1 · 1024 res, buckets 512–1536 · sigmoid timesteps, scale 1.3 · caption dropout 5%
Epoch 12 picked by comparing epochs 4–14 in 2:3, 16:9 and 21:9, with and without hires, 2 seeds each, image by image (epoch 10 hazier, epoch 14 smoother and less like the artist)