<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/b71b2474-ac3d-4b2c-a616-75256cc67ef4/original=true/b71b2474-ac3d-4b2c-a616-75256cc67ef4.jpeg" />Join The Tinkerer on Whop. Membership gets you early releases, private tools and a bunch of extra stuff.👉 <a target="_blank" rel="ugc" href="https://whop.com/cyberdelia-ai-lab">Join on Whop</a> 💬 <a target="_blank" rel="ugc" href="https://discord.gg/88R2hdjb22">Join the community for support, free tools and early news on Discord</a><hr />CyberRealistic Ernie-Image Turbo based on Baidu ERNIE-Image Turbo - is a distilled image model optimized for high-speed generation, capable of producing strong results in as little as 8 steps while maintaining impressive visual quality and prompt accuracy.Key features include excellent text rendering inside images, outperforming many competing models on benchmarks such as LongTextBench, along with bilingual prompt support for both English and Chinese. The model also utilizes the Flux 2 VAE to enhance overall image clarity, detail, and texture quality.CyberRealistic Ernie Turbo is especially well-suited for fast iteration, stylized visuals, poster/text-heavy compositions, and general-purpose image generation workflows.<hr /><h3 id="what-this-is-not">What this is not</h3>❌ Not trained on alien technology recovered from Area 51 ❌ Not guaranteed to survive “masterpiece, ultra detailed, 8k, cinematic” spam ❌ It will not fix your anatomy addiction ❌ It cannot read your mind when your prompt says only “girl standing” ❌ It won’t increase your GPU VRAM by sheer determination ❌ It is not powered by dark magic, quantum computing, or caffeine ❌ It may occasionally generate hands forged in another dimension ❌ Not responsible for sudden prompt addiction ❌ Side effects may include generating 400 test images at 3AM<h2 id="so...-what-is-this-then">So… what is this then?</h2>✅ A weird little experimental CyberRealistic project ✅ Fast enough to make rapid testing actually fun ✅ Built for curiosity, experimentation, and chaos ✅ Surprisingly capable for such a lightweight setup ✅ Great for trying ideas without waiting forever ✅ Released because experiments are more fun when shared CyberRealistic Ernie Turbo - ComfyUI Workflow V1 <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/9f641dd0-482b-46e2-83d1-361e8766216f/original=true/9f641dd0-482b-46e2-83d1-361e8766216f.jpeg" />A ready-to-run workflow for CyberRealistic Ernie, built around fast 8-step generation with an integrated AI upscale pipeline. The workflow file is in the Required Components section. <span style="background-color:rgb(26, 27, 30);color:rgb(141, 141, 143);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">Future updates and beta versions are available via the public Whop page: <a target="_blank" rel="ugc" href="https://whop.com/joined/cyberdelia-ai-lab/comfy-ui-workflows-ernie-9jzOtDw743Y8pq/app/">https://whop.com/joined/cyberdelia-ai-lab/comfy-ui-workflows-ernie-9jzOtDw743Y8pq/app/</a>

CyberRealistic Ernie-Image Turbo

Join The Tinkerer on Whop. Membership gets you early releases, private tools and a bunch of extra stuff. 👉 <a target="_blank" rel="ugc" href="https://whop.com/joined/cyberdelia-ai-lab/products/the-tinkerer/">Join on Whop</a>First CyberRealistic Ernie Turbo release, including a ready-to-use workflow for quick setup and experimentation.

cyberrealisticErnie_v10_txt.safetensors

flux2-vae.safetensors

cyberrealisticErnie_v10.safetensors

cyberrealisticErnie_v10.zip

v1.0

<a target="_blank" rel="ugc" href="https://huggingface.co/baidu/ERNIE-Image">HF</a> | <a target="_blank" rel="ugc" href="https://civarchive.com/api/download/models/2439129?type=Model&amp;format=SafeTensor&amp;size=pruned&amp;fp=fp8">VAE</a> |<a target="_blank" rel="ugc" href="https://civarchive.com/api/download/models/2857928?type=Training%20Data">Basic workflow</a> in the zip file. Turbo CFG: 1 Steps: 8ERNIE-Image is an 8B parameter text-to-image model from Baidu.It excels at complex prompt understanding, clear text rendering inside images, and structured layouts like posters, comics, and multi-panel designs.Includes a built-in Prompt Enhancer that turns short inputs into detailed descriptions. Works great in English and Chinese.Apache 2.0 licenseThe model is mirrored here for convenience.

Ernie-Image

ernieImage_txtEncMistral33b.safetensors

txt_enc Mistral-3-3b

Turbo

BF16

<h1 id="kirazuri-(ernie)">Kirazuri (Ernie)</h1>Kirazuri (Ernie) is a full fine-tune of the ERNIE-Image model.<a target="_blank" rel="ugc" href="https://huggingface.co/baidu/ERNIE-Image">ERNIE-Image</a> is an 8B DiT text-to-image generation model developed by the ERNIE-Image team at Baidu.Version 0.1 (Latest) is trained on ~50,000 images at multiple resolutions in three total stages up to a maximum resolution of 1024^2 pixels.This fine-tune focuses on several goals:- Learn new concepts/styles/characters associated with tag-based prompting.- Enhance the model aesthetic guided by manually applied quality, aesthetic, and style tagging- Preserve the text-rendering capabilities in multiple languages (English, Chinese, Japanese)For more details, see the <a target="_blank" rel="ugc" href="https://github.com/motimalu/diffusion-training-configs/blob/main/diffusion-pipe/ernie/notes/ernie-kirazuri.md">Kirazuri (Ernie) Training Diary</a><h2 id="training-details-summary">Training Details Summary</h2>Trainer: <a target="_blank" rel="ugc" href="https://github.com/tdrussell/diffusion-pipe">diffusion-pipe</a>Training device: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation EditionTotal training time: ~500 hrs (~20 days)Total samples seen (un-batched steps): ~1,328,000Training resolutions: 512^2, 768^2, 1024^2<h3 id="additional-features">Additional Features</h3>- Tag Dropout: 10% with protected first 8 tags- Tag Shuffle: Applied to last unprotected tags- Natural Language: 6 total Short and Long Caption variants<h2 id="installing-and-running">Installing and running</h2>Workflow:<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/4a2f1f3f-a72c-4532-84f8-61345b0eeee9/original=true/4a2f1f3f-a72c-4532-84f8-61345b0eeee9.jpeg" /><span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">The model is natively supported in ComfyUI. The above image contains a workflow; you can open it in ComfyUI or drag-and-drop to get the workflow.<h3 id="quantizations">Quantizations</h3>Int8 convrot quants of the model and text encoder are available on the model card.Generation speeds *(on NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition):BF16 Model and TE:~23gb vram, 1.35it/sINT8 Model and TE: ~13gb vram, 2.32it/s<h3 id="generation-settings">Generation Settings</h3>Recommendations:<ul><li>1024^2 Resolutions</li><li>50 Steps</li><li>CFG 4</li><li>er_sde or euler Sampler</li></ul>*Some previews are generated at 1280^2 resolutions - despite no training yet at this resolution, the model can still perform well for small details here.*The model can converge in as few as 20 steps, but anatomy and prompt adherence will suffer.<h3 id="prompting">Prompting</h3>This model is trained on combinations of booru-style tags and natural language captions.It does not perform well yet with tag-only prompts, and they are best used in combination with long natural language descriptions.This follows from the base models natural language only training, and intended use with a prompt enhancer that expands descriptions to long-format text.Tag orderTags combined with natural language can be placed at the start or end of the prompt.<code>[quality/meta/safety tags] [character] [series] [artist] [1girl/1boy/1other etc] [general tags]</code><code>[quality/meta/safety tags] [character] [series] [artist]</code> tag groups are also not shuffled, so their order may have some influence on generations.Quality and Aesthetic tagsHuman score based: <code>masterpiece, best quality, very aesthetic, aesthetic</code>Meta tagsabsurdres, official art, etcStylespainterly, chiaroscuro, ligne claire, flat color, no lineart, blending, etctraditional media, oil painting \(medium\), watercolor \(medium\), etc<h2 id="recognitions">Recognitions</h2><ul><li>Thanks to Baidu labs for the open research and release of the ERNIE-Image model.</li><li>Thanks to tdrussell of CircleStone Labs for the <a target="_blank" rel="ugc" href="https://github.com/tdrussell/diffusion-pipe">diffusion-pipe</a> trainer.</li><li>Thanks to narugo1992 and the deepghs team for open-sourcing various training sets, image processing tools, and models.</li><li>Thanks to silveroxides for the <a target="_blank" rel="ugc" href="https://github.com/silveroxides/convert_to_quant">convert_to_quant</a> quantization tool.</li></ul><h2 id="license">License</h2>Apache license 2.0Note that ERNIE-Image is governed by the same license; this fine-tune does not grant any rights or impose any restrictions to the base weights.

Kirazuri (Ernie)

kirazuriErnie_v01_txt_int8.safetensors

kirazuriErnie_v01_int8.safetensors

kirazuriErnie_v01_bf16.safetensors

kirazuriErnie_v01_txt_bf16.safetensors

v0.1

<a target="_blank" rel="ugc" href="https://huggingface.co/baidu/ERNIE-Image">HF</a> | <a target="_blank" rel="ugc" href="https://civitai.com/api/download/models/2439129?type=Model&amp;format=SafeTensor&amp;size=pruned&amp;fp=fp8">VAE</a> |<a target="_blank" rel="ugc" href="https://civitai.com/api/download/models/2857928?type=Training%20Data">Basic workflow</a> in the zip file. Turbo CFG: 1 Steps: 8ERNIE-Image is an 8B parameter text-to-image model from Baidu.It excels at complex prompt understanding, clear text rendering inside images, and structured layouts like posters, comics, and multi-panel designs.Includes a built-in Prompt Enhancer that turns short inputs into detailed descriptions. Works great in English and Chinese.Apache 2.0 licenseThe model is mirrored here for convenience.

Files with SHA256: 49a750a128863854eac7d85e1a277a7b44bf6ec3646405b84686dfeeca3708ca