## I. IntroductionMageTrail is a Danbooru/E621 proof of concept Full-Finetune of Microsoft's MageFlow 4B T2I model, using a diversity maximized condensed 41k images dataset as a way to tune booru concept and tags based prompting + Illustration capabilities into the model without having to tune with the full booru dataset. (which would have cost 50k+ dollars, no thanks).Starting from V0.25 onward, the model is now a completely new architecture due to being trimmed down to 2.8 billion parameters. This help the model become near SDXL/Anima range of availability for all users, with faster gen time and lower vram usage for finetuning/LoRA training. The trimming was done by <a target="_blank" rel="ugc" href="https://huggingface.co/Bluvoll/MageTrail-Shared-AdaLN-SVD">Bluvoll/MageTrail-Shared-AdaLN-SVD.</a>Please scroll down to the generation recommendation section to download the required custom node.The model is now also supported for easy consumer LoRA training/Finetuning through <a target="_blank" rel="ugc" href="https://github.com/bluvoll/mage-flow-trainer">bluvoll/mage-flow-trainer: A trainer for Mage-Flow with SDNQ support.</a>My second attempt at fine-tuning an image model on larger scale, this finetune aim to prove to the open source community on MageFlow 4B having good potential as an architecture for further finetuning.~ While V0.1 and V0.2-V0.25 is still very obviously undertrained and unstable (only 100 + 143 dollars spent, it's a minor miracle that it's learning this well), the model has shown great promise in quickly learning and adapting booru concept and tags to its knowledge base~ The architecture behind MageFlow 4B shows good promise for further investment:* Being 15-20% faster than NVIDIA Cosmos2/Anima on inference despite being 2 billion parameters larger* Using MageVAE which perform better than QwenVAE on all usage, only behind the strongest open source VAE currently being Flux2VAE (which it was distilled from), also having the ability to slot in Flux2VAE for inference* Having a decent Qwen 3 VL 4B Text Encoder* 256-2048 resolution native support* Being fairly quick to learn and adapt to new knowledge without any knowledge forgettingFuture goal for the project: gather funding of 1000~ dollars to finetune various artist styles into the model (planned V0.5-V1) and help further with convergence/aesthetic improvement.Any donation will help with achieving this goal, you can do so through:Crypto (Prefered, cause Kofi/Paypal money transfer time is ass and they take a big cut)0x6a4bc748cd0bb9ced9a360eb0eb79f4f106614f8 (USDT - BEP20 Network)12PPVYUeS1MerNp38Tpns5qXR6cmhu9tws (Bitcoin - BTC Network)0x6a4bc748cd0bb9ced9a360eb0eb79f4f106614f8 (Ethereum - ERC20 Network)FitfJAsxLUBuSgDJJaHgBXJpt1sMm5FzF1Tvf1SHW5Up (Solana - SOL network)Please handle your money carefully and make sure the address you're sending to is correct.Ko-fi<a target="_blank" rel="ugc" href="https://ko-fi.com/talanartvn**">https://ko-fi.com/talanartvn</a>## II. Model DetailsBase model: <a target="_blank" rel="ugc" href="https://huggingface.co/mage-flow-community/Mage-Flow">Microsoft's MageFlow 4B T2I</a>Method: Full-FinetuneTrainer: <a target="_blank" rel="ugc" href="https://github.com/RicemanT/diffusion-pipe-mageflow-ft">https://github.com/RicemanT/diffusion-pipe-mageflow-ft</a>Hardware: x8 H100 HBM3 80GB, rented with Banodoco sponsor and donation from the communityTotal training time: 64 (v0.1) + 92 (v0.2) H100 hoursTotal samples seen: 1247940 (v0.1) + 2911860 (v0.2)Training resolutions : 1024²## Training run## Version 0.1 (initial 20 epoch run → extended 10 epoch run)Budget: 130~ dollars (25-30 lost due to experiments and mistakes)Full config: <a target="_blank" rel="ugc" href="https://github.com/RicemanT/model-training-configs/blob/main/diffusion-pipe/MageFlow/configs/V0.1/mage_flow_BooruEssenceFFT.toml">Training</a> and <a target="_blank" rel="ugc" href="https://github.com/RicemanT/model-training-configs/blob/main/diffusion-pipe/MageFlow/configs/V0.1/mage_flow_BooruEssenceDataset.toml">Dataset</a>- Learning rate: 7e-6- LR scheduler: Warmup -&gt; Constant -&gt; REX to 0e-7- Precision: Full BF16- Optimizer: AdamW8bit with Kahan summation (to offset BF16 precision roundoff)- Weight decay: 0.02- Timestep sampling: Logit-Normal, shift 6, sigmoid scale 1.0## Version 0.2 (70 epoch continuation)Budget: 143~ dollarsFull config: Same as v0.1, just with changed lr due to only using x4 gpu instead of x8- Learning rate: 5e-6- LR scheduler: Warmup -&gt; Constant -&gt; REX to 0e-7- Precision: Full BF16- Optimizer: AdamW8bit with Kahan summation (to offset BF16 precision roundoff)- Weight decay: 0.02- Timestep sampling: Logit-Normal, shift 6, sigmoid scale 1.0## Additional training features- Tag dropout: 10%- Caption dropout: 5%- Mixed captions at 25/25/25/25 ratio (tags only, NL only, tags-nl, nl-tags)- Tag shuffle- Caption shuffle- Artist trigger attribution system## III. Recommended Settings<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/d5cd80cc-0d0c-48bf-b49e-1fbed7625e25/original=true/d5cd80cc-0d0c-48bf-b49e-1fbed7625e25.jpeg" />From V0.25 onwards, the model require downloading the <a target="_blank" rel="ugc" href="https://github.com/bluvoll/ComfyUI-MageFlow-Compressed">bluvoll/ComfyUI-MageFlow-Compressed custom node. </a>The model is not supported in any other UI currently. These are the settings used for the sample images above (ComfyUI):- Shift: 5.0 (or ComfyUI default)- Steps: 30- CFG: 5- Sampler: <code>euler_ancestral</code> or <code>er_sde</code>- Scheduler: <code>simple</code> or <code>beta</code>These are just my usual settings and workflow — feel free to experiment.Artist Trigger: This model use the Drawn by artistname trigger, if you want to use the model built-in artist tag, please always put one at the start of your prompt. (Currently V0.1 barely support any artist or characters though)Tagging System: This model support both danbooru and e621 tag prompting, although currently due to the model being severely undertrained alot of tags and concept are not working properly yet. Tags only and NL only prompt both work, but the model perform better with mixed prompting and clearer/slightly longer prompt.Example prompt: Drawn by nyatcha, m200 (girls' frontline), 1girl, :t, backpack, bag, binoculars, building, city, crane (machine), crossed bangs, crossed legs, feet out of frame, grey jacket, headset, holding, holding binoculars, jacket, long sleeves, messy hair, pout, short hair, sidelocks, sitting, solo, star (symbol). Set in a panoramic, dense futuristic cityscape during dusk, M200 from Girls' Frontline sits perched near an elevated railing on the right side of the frame with her legs crossed. She has short, messy grey hair with crossed bangs and sidelocks, wearing an olive-grey communication headset over her ears. She is dressed in a bulky, layered dark grey military jacket with long sleeves and carries a tactical backpack with mechanical gear protruding over her shoulder, resting both hands in front of her while holding a pair of compact black binoculars. The bustling urban backdrop features towering, dark multi-story architecture filled with glowing square windows, exposed steel framing, rooftop antennas, and a large red construction crane against a hazy, overexposed bright sky. Cars with gleaming roofs pass below the railing amidst the muted industrial color palette of beige, black, and amber lighting.## IV. Dataset<a target="_blank" rel="ugc" href="https://huggingface.co/datasets/RicemanT/booru-essence-2026">Booru-Essence-2026-41k images</a>Originally created by Lodestone Rock, the dataset was updated to 2026 tag standard and captioned with SOTA API captioners, see dataset repo for details.Model training, Dataset and Captioning tooling lives in the <code>utils</code> folder](<a target="_blank" rel="ugc" href="https://github.com/RicemanT/model-training-configs/tree/main/diffusion-pipe/MageFlow/utils">https://github.com/RicemanT/model-training-configs/tree/main/diffusion-pipe/MageFlow/utils</a>) of the training repo.---## VI. Notes from the Training DiaryFull training diary: <a target="_blank" rel="ugc" href="https://github.com/RicemanT/model-training-configs/blob/main/diffusion-pipe/MageFlow/notes/MageTrail-diary.md">MageTrail-diary</a>---## VII. LicenseThis model is a Derivative of Microsoft's MageFlow and is distributed under the same [MIT License](<a target="_blank" rel="ugc" href="https://choosealicense.com/licenses/mit/)**">https://choosealicense.com/licenses/mit/)</a> as the base model, with no additional restrictions.---## VIII. AcknowledgmentsBeeg thanks to:- [Banodoco](<a target="_blank" rel="ugc" href="https://www.banodoco.ai/">https://www.banodoco.ai/</a>) and [their Discord](<a target="_blank" rel="ugc" href="https://discord.gg/yzwcNaSEz">https://discord.gg/yzwcNaSEz</a>) — Their 88.77 dollar grant made this project possible, the biggest thanks to them- [Lodestone Rock](<a target="_blank" rel="ugc" href="https://huggingface.co/lodestones">https://huggingface.co/lodestones</a>) — Creator of the original version of the dataset that this model is trained on- [Motimalu](<a target="_blank" rel="ugc" href="https://civarchive.com/user/motimalu">https://civarchive.com/user/motimalu</a>) — Inspiration behind finetuning practices and configs- [Bluvoll](<a target="_blank" rel="ugc" href="https://github.com/bluvoll/diffusion-pipe">https://github.com/bluvoll/diffusion-pipe</a>) — diffusion-pipe fork derived from to use for training, and general training advice- [Anzhc](<a target="_blank" rel="ugc" href="https://huggingface.co/Anzhc">https://huggingface.co/Anzhc</a>) — general training advice- [Nruaif](<a target="_blank" rel="ugc" href="https://huggingface.co/Shio-Koube">https://huggingface.co/Shio-Koube</a>) — diffusion-pipe fork derived from to use for training, and general dataset handling/training advice- [Astromahdi](<a target="_blank" rel="ugc" href="https://gpu.garden/">https://gpu.garden/</a>) — jupyter workspace where I processed and store the dataset- [animetimm/DeepGHS](<a target="_blank" rel="ugc" href="https://huggingface.co/animetimm/convnextv2_huge.dbv4-full">https://huggingface.co/animetimm/convnextv2_huge.dbv4-full</a>) — Danbooru tagging model- [RedRocket](<a target="_blank" rel="ugc" href="https://huggingface.co/RedRocket/Hydra">https://huggingface.co/RedRocket/Hydra</a>) — E621 tagging model

MageTrail (MageFlow 4B Dannboru/E621 Finetune)

This is not a further finetune, it's just a compressed version of V0.2. MageTrail from this point on is a completely new architecture seperate (but still based) on MageFlow 4B, due to approximating the AdaLN layers of the model with SVD at rank 256, basically getting near full AdaLN results for a fraction of parameters.The model is now a 2.8B architecture, in the same tier as SDXL and Anima, with even slightly faster generation speed and alot lower vram usage for inference and training compared to the 4B version. All thanks to the work of <a target="_blank" rel="ugc" href="https://huggingface.co/Bluvoll/MageTrail-Shared-AdaLN-SVD">Bluvoll/MageTrail-Shared-AdaLN-SVD</a>

Files with SHA256: 9096877a043d84ca57f2f00ef162430f1ce13b8915ddaa0075d1b892abae6285