Real Dream Website (Faster download)
Available at TensorArt: https://tensor.art/models/899250342552672189/
Comfyui Workflows:
Anima text2image with Loras Workflow
Anima img2img (denoise) with Loras Workflow
Klein 9B Edit with Loras Workflow
Klein 9B img2img (denoise) with Loras Workflow
Z-image-turbo img2img workflow
Z-image text2img with loras workflow
Z-image De-Turbo with loras workflow
This model is capable of generating very realistic images that can confuse people's critical sense, please do not use it irresponsibly, harming ordinary people or public figures. Also avoid its use in electoral campaigns, promoting or damaging the image of candidates. Although the field of artificial intelligence is poorly regulated, its improper use can cause legal penalties even in countries with little regulation in the area. Be very careful to not commit crimes in your country and not to promote injustices against innocent people. I do not authorize the use of this model for Lora training using the face of real people without authorization from the biological owner. Do not create images that can embarrass and damage the public image of people. This can cause serious psychological harm, including deep depression. I do not authorize the use of images generated by this model for scam applications and the dissemination of false information.
During the fine-tuning of AI models, I only use synthetic images that do not infringe any copyright, personal image rights, or laws. The models are merges of open models from different sources and may contain diluted data remnants in the neural network. I am not responsible for possible data remnants since they are diluted, and it is not humanly possible to track the existence and origin of this diluted data.
I do not allow the upload of any of my models on Tensor Art. I only do not allow it there because the policy of duplicate models on Tensor Art prevents me from uploading again if someone has already done it.
I have enabled to automatically hide all NSFW posts for several reasons that are not worth listing. I hope you have patience and understand.
Description
FAQ
Comments (12)
Hello?
I am very sorry if this question is stupid, but can anyone explain, what is the difference between Klein 9B and Klein 9B Base? Please?
Distillation: "Flux2Klein 9B" is the distilled version. It has been optimized (via techniques like distillation) to produce high-quality images in very few steps (usually 4). This makes it much faster while still delivering excellent results for everyday text-to-image and image-editing tasks.
Base Model: "Klein 9B Base" is the undistilled foundation model. It preserves the full training signal and offers greater flexibility and control. You can run it with more steps for potentially higher quality in some cases, and it's especially suited for fine-tuning, training LoRAs, or advanced/custom workflows where you need maximum model capacity.
In practice:
Use Flux2Klein 9B (distilled) when you want speed without much quality loss — great for interactive apps, quick iterations, or local inference on consumer GPUs.
Use Klein 9B Base when you need maximum quality/control or plan to fine-tune the model yourself.
@ulfrathje That's a wonderful answer, thank you!
@Renesi Distilled models generally have less variety, which is a trade-off for being able to render a coherent image with fewer steps. Sometimes distillation (like Wan light2v/lightning loras) tends to exaggerate contrast and make the lighting less realistic.
@ulfrathje Shouldn't the 'base' model be 16gb? Why are they both the same size?
@xshan they are both fp8 format, which cuts the size roughly in half. in this case the model architectures for both are the same, so they are the same size. the only difference is that the non-base version has been trained to generate the image in less steps, but it loses some quality and prompt following ability due to that. it's the same difference between z-image and z-image turbo, klein just doesn't call it turbo
Sad that no BF16 (or even FP16) is available for ZIB. Wanted to give it a try til I noticed that. I feel like high vRAM users are being discriminated against.
@freestuffpl0x42069 Thanks
I have a problem with z image Wich is getting very darks ,dimly lit outputs
Keep the CFG between 1 and 3 and the steps between 8 and 20. If you use a lot of Loras with not enough steps, it generally helps to put CFG at least at 2 and move up the steps to 20. Just generate the same prompt twice, one with CFG 1 and steps 8, then CFG 2 or 3 and steps 20, and see if it gets better.
will the bf16 model come out? looking forward to
Post your images with meta data, ask for help provide link.