V1.9

(Turbo included: CFG 1.2, Steps 16, I like ER_SDE/simple)

Not yet the final form! This version is an upgrade of the Failed Experiment to get back on track with this model version. To do so, i leveraged Sam ANIMA Realistic Turbo V2.3. First i converted it to ANIMA 2.9B, then i did some level of merge into Failed Experiment:

>>> import sd_mecha as sdm
... a = sdm.model("UnNamedA29-FailedExperiment.safetensors")
... b = sdm.model("samANIMARealistic_turboV23-A29.safetensors")
... c = sdm.slerp(a,b,alpha=0.3)
... d = sdm.slerp(a,b,alpha=0.5)
... e = sdm.n_average(a,c,d)
... sdm.merge(e,output="alpha.safetensors",output_dtype="bfloat16")

After that, it was some touch and go with the various upgrader LoRAs i had made previously for UnNamedA 2B, converted to 2.9B (including one based on muKrea).

I also built a custom Qwen3-0.6B TE based on the base TE + my Qwen-Heretic LoRA (at 0.2) + DARE/TIES (at 0.3) from the uncensored TE by @Felldude (i didn't use it as-is as it lost a bit of ANIMA knowledge from my test). This TE (together with a VAE) is baked in the checkpoint and will not be published as a standalone.

V2 when i can set up a better training workflow!

Version comparison: Failed vs V1.9 vs source SAM model

Failed Experiment

12h (on 18) in my training of V1, i had a power outage... so i took a minute to check where i was. Epoch 9... of a failed training. Because, i started the training on the Turbo version of the model instead of the normal one. Let's say it messed up a few things 😅

Not wanting to throw it away, i managed to salvage stuff and here we are. An INT8 Convrot standalone version of this model. It kinda works, but not in the way I wanted. Works like V1, is not AIO but with Turbo. And gives better results with Euler A / Simple.

V1

This model is a finetuning of my previous conversion to ANIMA 2.9B of UnNamedA.

Usage:

• /!\ Turbo model: CFG 1~1.5, Steps 8~16

• ER SDE / Beta or euler A / sgm uniform

• Positive prompt: masterpiece, best quality, highres, realistic, <your prompt>

• Negative prompt: worst quality, low quality, blurry, anatomical nonsense, text, watermark, signature (optional: poorly drawn face, mutated face, ugly, cropped, out of frame, 3d, cgi, render)

• Flow Shift: 3

• HiRes.Fix: denoise 0.3, same CFG, same number of steps has standard pass, 4xNMKD Superscale

• AIO: This model contains the VAE and TE

Build:

Unlike classic ANIMA models, UnNamedA also has an heritage from Cosmos 2B which brought some "glitches" when converted (some strange change in the face and a "grain" on the pictures). That's why i de-turboed UnNamedA 2.0, converted it to 2.9B and did a new pass (two actually) of finetuning directly on the 2.9B version using this tool.

After some testing and merging to avoid outliers, i re-merged at 0.7 a new turbo (it really gives better results): the original cosmos2 turbo to avoid bringing anime aesthetic in the mix was converted into 2.9B via this script and then cleaned of incompatible keys.

Here is a quick comparison:

• UnNamedA V2.0t-29B

• UnNamedA29 V1

The "grain" can be seen all over the hair and skin on the previous version and the teeth where a bit strange.

NB: Yes, I'll do a V2 to further train this model, since i feel it lost a bit of "texture" in the process (and this time, i'll do a proper teacher/student training).

PS: A V1.5 is undergoing training:

GPU goes Brrrr (64 dim LoRA, 1536px resolution, i am nearing what i can train locally)

~8h done, ~10 to go!

Slowly going doooowwwnnnn...

Thanks for reading! 💜