New top local AI image generator is here! Already uncensored

Creal 2 is a newly released open-source image generator that stands out for being fully uncensored out of the box, with free downloadable weights that run offline for unlimited generations in about 10 seconds on consumer GPUs with just 8 GB of VRAM. The video provides a ComfyUI-based tutorial covering the necessary models (Turbo FP8 diffusion, Qwen 3 VL text encoder, VAE), key workflow settings like CFG and step counts, and the optional Conditioning Creal Rebalance node for robust spicy outputs. Beyond its explicit-content capability, the model excels at photorealism, anatomy, text rendering, and prompt adherence, which the presenter attributes partly to the fact that no AI-generated images were used in pretraining. The technical report also reveals that very-large-scale training runs were highly unstable and frequently crashed, a major scaling challenge. Licensing is permissive for a community model, allowing commercial use for companies earning under $1 million annually, making it a strong alternative to models like Ideogram or Flux.

• ▶ 0:00 Creal 2 is a newly released open-source image generator with free, downloadable weights, making it the "most uncensored" and able to run offline for unlimited generations.
• ▶ 0:14 It is fast and accessible, generating images in about 10 seconds on consumer GPUs with as little as 8 GB of VRAM.
• ▶ 1:39 Unlike other models (Ideogram, Z Image, Flux), Creal 2 can generate explicit/spicy content right out of the box—this capability is built into the base model, while also excelling at photorealism, anime, anatomy, world knowledge, and text rendering.
• ▶ 2:02 The tutorial relies on ComfyUI, which supports automatic offloading so large models can run on low VRAM.
• ▶ 3:13 In the Templates sidebar, select the fully offline "Krea text-to-image" template, not the API-based one.
• ▶ 4:18 Download the Krea 2 Turbo FP8 diffusion model, the Qwen 3 VL text encoder (~4.8 GB), and the VAE (~250 MB), placing each in the correct ComfyUI model folder.
• ▶ 5:21 Key workflow settings: prompt input, prompt enhance toggle (usually off for speed), image size/aspect ratio up to 21:9, megapixels up to 4K, seed, and LoRA disabled by default.
• ▶ 6:46 Important generation controls: 8 steps is usually enough for the turbo model, CFG controls prompt adherence (higher = more literal, lower = more creative), and sampler/scheduler determine the generation algorithm.
• ▶ 7:26 First run on an RTX 5080 (16 GB VRAM) took 14 seconds including model loading, showing the image generation speed for this setup.
• ▶ 10:48 LoRAs are added by selecting the model and setting strength; most require specific trigger words in the prompt to activate the style.
• ▶ 12:29 Creal 2 is uncensored out of the box, but for robust spicier outputs you need to install the Conditioning Creal Rebalance node and connect it between the prompt and K Sampler.
• ▶ 14:29 Summary: Creal 2 is one of the best open-source models, with excellent prompt adherence, anatomy, text rendering, and very fast generation (~10 seconds).
• ▶ 15:36 The technical report reveals that no AI-generated images were used in pretraining, which the presenter cites as a key reason for the model's strong aesthetics and realism.
• ▶ 17:09 Scaling training was a major challenge: training on 128 GPUs was stable, but larger runs were unstable and crashed so frequently that not a single very-large-scale run completed without a crash.
• ▶ 17:52 The licensing is very permissive for a community model, allowing commercial use if annual revenue is under $1 million USD, making it more flexible than models like Ideogram or Flux.
• ▶ 18:25 Wraps up the review and tutorial, inviting viewers to share their opinions in the comments.
• ▶ 18:31 Offers troubleshooting help for installation issues, asking viewers to copy and paste exact error messages in the video description.
• ▶ 18:50 Promotes the free weekly newsletter to stay up-to-date on AI news, alongside a call to like, share, and subscribe.

• ▶ 0:00 Introduction and Capabilities (0:00 - 2:02) - - Introduces Creal 2, an uncensored open-source image generator with photorealistic, top-notch aesthetics.
• ▶ 2:02 Setup and Model Download in ComfyUI (2:02 - 5:21) - - Covers ComfyUI installation, downloading Krea 2 and Creatus model files, and selecting the right models in the workflow.
• ▶ 5:21 Workflow Settings and First Run (5:21 - 10:00) - - Explains workflow settings, running the model on an RTX 5080, reusable workflows via Luma Skills, and a fast image generation speed test.
• ▶ 10:00 LoRA Use, Censorship, and Summary (10:00 - 14:54) - - Demonstrates adding and using LoRAs, managing uncensored output with a rebalance node, and recaps Creal 2's strengths.
• ▶ 14:54 Training, Technical Report, and Licensing (14:54 - 18:25) - - Covers LoRA training with the AI Toolkit, the technical report on data filtering and GPU scaling, and licensing terms.
• ▶ 18:25 Outro and Call to Action (18:25 - 19:11) - - Wraps up the review and asks viewers for their thoughts.