Winnougan Presents: 10 Toes INT8 Convrot and FP8
10 Toes V1.3:
What's changed?
Some users reported that V1.2 was too sanitized at times, and not as spicy as V1. V1, however, struggled at some SFW images and was throwing in naughty bits when not wanted.
I decided to go back to the kitchen and do some more cooking. After several tests, I finally came out with goose's golden egg. And here we are with files to satiate any gooner's wildest needs. We have:
BF16, FP8, INT8 Convrot and GGUF variants: Q8 and Q4.Updated workflow: Winnougan_10ToesV1.3_Krea_2_WF - Pastebin.com
The workflow includes BF16, FP8, INT8 Convrot or GGUF with a simple switch.
The workflow is dead simple with a little RTX Upscaler thrown in for good measure.
Need help? Join my Discord. Wanna support me? Ko-fi please. Links at bottom of this post.
10 Toes V1.2:
Winnougan 10 Toes V1.2 — Available Versions
Three versions are available depending on your hardware and priorities:
BF16 — the original, full precision. Best quality, highest VRAM requirement. Use this if you have the headroom.
FP8 — lighter quantization with minimal quality difference from BF16. Good starting point for most users, works on any recent GPU, fastest inference of the two quantized versions.
INT8 ConvRot — most aggressive VRAM reduction, using group-wise Hadamard rotation (group size 256) to suppress outliers before INT8 quantization. Best quality-to-size ratio for Ampere GPUs (RTX 3000 series and newer). Slightly slower than FP8 but noticeably more VRAM-efficient.
All quantized versions converted with convert_to_quant v1.3.1 on an RTX 4090.
Which should I use?
8GB VRAM or less → INT8 ConvRot
10–12GB VRAM → FP8
16GB+ VRAM → BF16
10 Toes' Genesis:
This model is proper filthy. Everything was made locally on my RTX 4090. I've been making Krea 2 LoRAs since day 1 of the model's launch in AI Tookit. I've made a few tutorials on the matter on YouTube: Winnougan - YouTube
I made a few secret recipes NSFW LoRAs that I never released with over 5k high resolution images in my dataset. I trained that LoRA with the help of all the NSFW LoRAs from Civitai Red in Krea 2. I love all the NSFW LoRAs out there - they're all awesome. With my new LoRAs made, I burnt them into the DNA of the BF16 Krea 2 Tubo checkpoint. My jaw hit the floor like Roger Rabbit catching a glimpse at Jessica Rabbit for the first time. The images had perfect glowing skin. All the mature prompts came through with flying colors.
I'm proud to present this model for you guys to goon to. As always, bring your soft Egyptian cotton sock and a jar of mayonnaise or a well-lubed fleshlight and get ready.
Since this is a Krea 2 Turbo checkpoint, I recommend the following settings:
CFG 1, 8-9 steps, Scheduler: Euler/Euler Ancestral or Er_SDE, Sampler: Beta/SimpleI've included a link to my workflow. If you have questions come to my Discord - it's full of gooners and AI bros like you.
🚨Important🚨:
There are two versions to download the INT8 Convrot (I set the listing to "MXFP8" because Civitai doesn't have an "INT8 Convrot" option) and the FP8 version. Both have been tested till their balls drained flaccid in ComfyUI. I have no idea outside of ComfyUI.
What the fuck is "INT8 Convrot" bro?
Good question - it's an optimized INT8 quantization that let's you goon hard and fast in ComfyUI. If you're running the nightly build of Comfy, with Pytorch 2.10 or higher, Python 3.12 or higher and cu130 or higher, then you're golden. It adds significant speeds to your renders. I've also uploaded on Huggingface the text encoders in INT8 Convrot too!
Consider supporting me on Ko-fi. Even $3 goes a long way!
I'm burning through resources, electricity bills and shaking with fear if anything should happen to my ram sticks or SSDs in this alternate reality where ram costs more than a G-Wagon.
👇🏻👇🏾Support me below 👇🏿👇🏼
💚 Workflow: (https://pastebin.com/nHWpSccV)
💙 Support on Ko-fi: (https://ko-fi.com/winnougan)
I'm saving toward a hardware upgrade so I can expand into Ideogram 4 and LTX-2.3 video LoRAs and deliver releases faster. Every coffee helps make it happen. ☕
All LoRAs have the last 4-8 epochs available for members on Ko-fi along with datasets and sample training images.
🖤 YouTube — Tutorials and reviews: https://www.youtube.com/@Winnougan
💀 Discord — Releases, requests, and conversation: https://discord.gg/CJv5wceJaN
Description
Winnougan 10 Toes V1.2 — INT8 Dynamic ConvRot
This is an INT8 quantized version of the Winnougan 10 Toes V1.2 Krea 2 Turbo finetune, converted using Dynamic ConvRot — a group-wise Hadamard rotation technique that automatically selects the optimal rotation block size per layer (minimum 256, scaling up where the architecture allows) before INT8 quantization. This produces sharper, more faithful output compared to plain INT8 by suppressing the row-wise outliers characteristic of DiT architectures before quantization, rather than letting them degrade the quantization scale.
The --krea2 preset keeps the first layer, last layer, text fusion projector, modulation layers, and other precision-sensitive components at native BF16, while quantizing the core transformer blocks. The result is a meaningfully smaller checkpoint that loads and runs within 8GB VRAM with no perceptible quality loss compared to the BF16 original.
Converted with silveroxides/convert_to_quant v1.3.1 on an RTX 4090.