<div>
<p><strong>mhDreamy</strong> is a <strong>dreamy cinematic</strong> 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.</p>
<p>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.</p>
<p>This is a <strong>style</strong> LoRA — not a character, not a concept. Trained entirely on the author's own material; no third-party copyrighted work was used.</p>
</div>

<h3>⚡ Recommended weight (read first)</h3>
<p>Training used <code>alpha/rank = 16/32</code> (half strength), so the raw file wants a higher slider value:</p>
<ul>
<li><strong>Krea 2 Turbo / quantized checkpoints</strong> (int8, nf4, fp8…): use <strong>2.0</strong> — 4-bit quantization swallows part of the LoRA delta.</li>
<li><strong>Official raw bf16 checkpoint</strong>: use <strong>1.0 – 1.5</strong>.</li>
</ul>

<h3>🔑 Trigger words</h3>
<pre>mhDreamy, dreamy golden-lit ethereal illustration</pre>
<p>Training caption format (content first, style anchor last):</p>
<pre>&lt;your content description in English&gt;. mhDreamy, dreamy golden-lit ethereal illustration</pre>
<p>Example prompts:</p>
<pre>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</pre>
<pre>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</pre>

<h3>✅ Where it works</h3>
<ul>
<li>Same-tonality cinematic illustration: characters, interiors, landscapes, animals. Fog, backlight, night rain, candlelit rooms and golden-hour exteriors are its home ground.</li>
<li>Measured style-fingerprint score (VGG19 Gram vs. training set): <strong>0.816 base model → 0.891 round 1 → 0.901 this version</strong>. On subjects that never appear in the training set it still lifted the score by <strong>+0.036 ~ +0.049</strong> — it learned the tonality, not the training images.</li>
</ul>

<h3>⚠️ Honest limitations</h3>
<ul>
<li><strong>Anime drifts.</strong> 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.</li>
<li><strong>Trigger-word leakage.</strong> When the prompt contains <code>mhDreamy</code> and the scene has a natural text surface (shop signs, posters), the trigger word can be painted into the picture. Add <code>text, watermark, signboard, letters, signature</code> to your negative prompt.</li>
<li>Hand/prop junctions (nets, hilts, instruments) occasionally melt — normal for this class of model.</li>
<li>Out-of-domain subjects (modern streets, wide cityscapes) barely move: the style is tonality-specific.</li>
</ul>

<h3>🧪 Training details</h3>

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 · sha256 <code>d27cb98383627c58…</code>


<h3>🖼 About the example images</h3>
<p>All examples are the LoRA applied to subjects that are <em>not</em> 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.</p>

<h3>📄 License</h3>
<p><strong>RentCivit</strong> — 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.</p>

mhDreamy — Dreamy Cinematic Style (Krea 2)

<div>
<p><strong>mhDreamy</strong> is a <strong>dreamy cinematic</strong> 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.</p>
<p>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.</p>
<p>This is a <strong>style</strong> LoRA — not a character, not a concept. Trained entirely on the author's own material; no third-party copyrighted work was used.</p>
</div>

⚡ Recommended weight (read first)
<p>Training used <code>alpha/rank = 16/32</code> (half strength), so the raw file wants a higher slider value:</p>
<ul>
<li><strong>Krea 2 Turbo / quantized checkpoints</strong> (int8, nf4, fp8…): use <strong>2.0</strong> — 4-bit quantization swallows part of the LoRA delta.</li>
<li><strong>Official raw bf16 checkpoint</strong>: use <strong>1.0 – 1.5</strong>.</li>
</ul>

🔑 Trigger words
<pre>mhDreamy, dreamy golden-lit ethereal illustration</pre>
<p>Training caption format (content first, style anchor last):</p>
<pre>&lt;your content description in English&gt;. mhDreamy, dreamy golden-lit ethereal illustration</pre>
<p>Example prompts:</p>
<pre>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</pre>
<pre>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</pre>

✅ Where it works
<ul>
<li>Same-tonality cinematic illustration: characters, interiors, landscapes, animals. Fog, backlight, night rain, candlelit rooms and golden-hour exteriors are its home ground.</li>
<li>Measured style-fingerprint score (VGG19 Gram vs. training set): <strong>0.816 base model → 0.891 round 1 → 0.901 this version</strong>. On subjects that never appear in the training set it still lifted the score by <strong>+0.036 ~ +0.049</strong> — it learned the tonality, not the training images.</li>
</ul>

⚠️ Honest limitations
<ul>
<li><strong>Anime drifts.</strong> 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.</li>
<li><strong>Trigger-word leakage.</strong> When the prompt contains <code>mhDreamy</code> and the scene has a natural text surface (shop signs, posters), the trigger word can be painted into the picture. Add <code>text, watermark, signboard, letters, signature</code> to your negative prompt.</li>
<li>Hand/prop junctions (nets, hilts, instruments) occasionally melt — normal for this class of model.</li>
<li>Out-of-domain subjects (modern streets, wide cityscapes) barely move: the style is tonality-specific.</li>
</ul>

🧪 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 · sha256 <code>d27cb98383627c58…</code>


🖼 About the example images
<p>All examples are the LoRA applied to subjects that are <em>not</em> 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.</p>

📄 License
<p><strong>RentCivit</strong> — 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.</p>

Files with SHA256: d27cb98383627c58b20b48004dfaa60edb300bc6b27839c33ff3051278987f4f