mhDreamy is a dreamy cinematic illustration style — soft anamorphic-looking haze, golden practical lights blooming through fog, deep contrast with gentle highlight roll-off, and a painterly film-grain finish. It reads like a still frame from a dreamy arthouse film rather than a flat digital illustration.
Core look: warm gold-and-amber key light, cool shadow separation, visible atmosphere (mist / dust / drifting light), rim-lit silhouettes, shallow-feeling depth and softly compressed backgrounds.
This is a style LoRA — not a character, not a concept. Trained entirely on the author's own material; no third-party copyrighted work was used.
⚡ Recommended weight (read first)
Training used alpha/rank = 16/32 (half strength), so the raw file wants a higher slider value:
- Krea 2 Turbo / quantized checkpoints (int8, nf4, fp8…): use 2.0 — 4-bit quantization swallows part of the LoRA delta.
- Official raw bf16 checkpoint: use 1.0 – 1.5.
🔑 Trigger words
mhDreamy, dreamy golden-lit ethereal illustration
Training caption format (content first, style anchor last):
<your content description in English>. mhDreamy, dreamy golden-lit ethereal illustration
Example prompts:
An elderly fisherman with a white beard mends a torn net on the deck of a small wooden boat while a lantern burns beside his knee. Thick fog drifts over the dark water behind him in a wide shot. mhDreamy, dreamy golden-lit ethereal illustration
A violinist in a dark coat plays alone on an old stone bridge in heavy rain, water streaming from the arch into the river below. Street lamps glow through the downpour in a wide shot. mhDreamy, dreamy golden-lit ethereal illustration
✅ Where it works
- Same-tonality cinematic illustration: characters, interiors, landscapes, animals. Fog, backlight, night rain, candlelit rooms and golden-hour exteriors are its home ground.
- Measured style-fingerprint score (VGG19 Gram vs. training set): 0.816 base model → 0.891 round 1 → 0.901 this version. On subjects that never appear in the training set it still lifted the score by +0.036 ~ +0.049 — it learned the tonality, not the training images.
⚠️ Honest limitations
- Anime drifts. On clean cel-shaded anime it pushes toward thick-paint / semi-realistic rendering. For anime, drop below 0.6 — or skip the LoRA and use the prompt anchor alone.
- Trigger-word leakage. When the prompt contains
mhDreamyand the scene has a natural text surface (shop signs, posters), the trigger word can be painted into the picture. Addtext, watermark, signboard, letters, signatureto your negative prompt. - Hand/prop junctions (nets, hilts, instruments) occasionally melt — normal for this class of model.
- Out-of-domain subjects (modern streets, wide cityscapes) barely move: the style is tonality-specific.
🧪 Training details
Base modelKrea 2 (official raw bf16) Dataset75 images, own material, 1024px Networkrank 32 / alpha 16 (scale 0.5) Round 12000 steps, lr 1e-4, from scratch Round 2 (this file)1500 steps, lr 5e-5, resumed from round 1 Optimizeradamw8bit · batch 1 · ai-toolkit 0.13.19 Filebf16 safetensors · 218 MB · sha256d27cb98383627c58…
🖼 About the example images
All examples are the LoRA applied to subjects that are not in the training set (fisherman at night, knight in snow, scholar in a candlelit library, violinist in the rain, astronomer at his window), plus a few training-set subjects. Same seed per subject, LoRA weight 1.0, 1024×1536, 25 steps, CFG 4.0.
📄 License
RentCivit — free to use, merge and share derived models; commercial use allowed. Credit appreciated but not required. Please do not re-upload the raw file as-is.
Description
mhDreamy is a dreamy cinematic illustration style — soft anamorphic-looking haze, golden practical lights blooming through fog, deep contrast with gentle highlight roll-off, and a painterly film-grain finish. It reads like a still frame from a dreamy arthouse film rather than a flat digital illustration.
Core look: warm gold-and-amber key light, cool shadow separation, visible atmosphere (mist / dust / drifting light), rim-lit silhouettes, shallow-feeling depth and softly compressed backgrounds.
This is a style LoRA — not a character, not a concept. Trained entirely on the author's own material; no third-party copyrighted work was used.
Training used alpha/rank = 16/32 (half strength), so the raw file wants a higher slider value:
- Krea 2 Turbo / quantized checkpoints (int8, nf4, fp8…): use 2.0 — 4-bit quantization swallows part of the LoRA delta.
- Official raw bf16 checkpoint: use 1.0 – 1.5.
mhDreamy, dreamy golden-lit ethereal illustration
Training caption format (content first, style anchor last):
<your content description in English>. mhDreamy, dreamy golden-lit ethereal illustration
Example prompts:
An elderly fisherman with a white beard mends a torn net on the deck of a small wooden boat while a lantern burns beside his knee. Thick fog drifts over the dark water behind him in a wide shot. mhDreamy, dreamy golden-lit ethereal illustration
A violinist in a dark coat plays alone on an old stone bridge in heavy rain, water streaming from the arch into the river below. Street lamps glow through the downpour in a wide shot. mhDreamy, dreamy golden-lit ethereal illustration✅ Where it works
- Same-tonality cinematic illustration: characters, interiors, landscapes, animals. Fog, backlight, night rain, candlelit rooms and golden-hour exteriors are its home ground.
- Measured style-fingerprint score (VGG19 Gram vs. training set): 0.816 base model → 0.891 round 1 → 0.901 this version. On subjects that never appear in the training set it still lifted the score by +0.036 ~ +0.049 — it learned the tonality, not the training images.
- Anime drifts. On clean cel-shaded anime it pushes toward thick-paint / semi-realistic rendering. For anime, drop below 0.6 — or skip the LoRA and use the prompt anchor alone.
- Trigger-word leakage. When the prompt contains
mhDreamyand the scene has a natural text surface (shop signs, posters), the trigger word can be painted into the picture. Addtext, watermark, signboard, letters, signatureto your negative prompt. - Hand/prop junctions (nets, hilts, instruments) occasionally melt — normal for this class of model.
- Out-of-domain subjects (modern streets, wide cityscapes) barely move: the style is tonality-specific.
d27cb98383627c58…
🖼 About the example images
All examples are the LoRA applied to subjects that are not in the training set (fisherman at night, knight in snow, scholar in a candlelit library, violinist in the rain, astronomer at his window), plus a few training-set subjects. Same seed per subject, LoRA weight 1.0, 1024×1536, 25 steps, CFG 4.0.
📄 LicenseRentCivit — free to use, merge and share derived models; commercial use allowed. Credit appreciated but not required. Please do not re-upload the raw file as-is.