Flux 2
The speaker argues that Flux 2, a 32B guidance-distilled model that does both text-to-image and image editing, can be trained with AI Toolkit on day zero. A style LoRA trained with a trigger word and Differential Output Preservation (DOP) can learn a painter's style while keeping the model's broader prompt and image knowledge, and the result can then be invoked on edit inputs. The cost is very high VRAM, and the speaker presents the approach as an experiment.
The speaker frames Flux 2 as the successor to Flux 1 and the editing models (Kontext, Qwen). It is one 32B model that handles both text-to-image and editing, and a 24B vision-language text encoder and a new 32-channel autoencoder come with it. The speaker got early access from Black Forest Labs so AI Toolkit could support it on release day.
Size is presented as the central practical obstacle. The speaker calls the model 'very big' and says that even a 96 GB GPU needs 8-bit quantization. Training needs about 64 GB of VRAM and does not fit on 48 GB, so consumer hardware looks unlikely at launch. The speaker still intends to push the floor lower, citing past cases where 'it can't be done' and was done anyway.
Because one model does both generation and editing, the speaker argues a style or character can be taught on plain text-to-image data. It can then be applied to a supplied picture with an instruction such as 'make this picture [trigger] style'. This avoids building paired edit datasets or inpainting characters onto other people.
How the data is captioned decides what the LoRA learns. The speaker says the trigger word absorbs everything the captions leave out, so captions should describe content and omit what should become the style.
The sample results are offered as evidence that style and general understanding can coexist. The style arrives quickly, complex scenes stay coherent, and edit inputs such as a cat and a Mona Lisa are restyled with fine paint-stroke detail. The speaker credits the preservation approach for the retained comprehension and the new 32-channel autoencoder for the fine detail.
The speaker reports flaws and unknowns openly. A biased dataset turned a male subject into a woman and removed the speaker's dad's mustache. The speaker is unsure whether instruction-style training is the right way to do this, the VRAM floor is unknown, and ComfyUI support is unconfirmed at recording time. Follow-ups are planned on character swapping and lower-VRAM training.
Guidance-distilled training at guidance 1 — Flux 2 is guidance-distilled like Flux 1 dev, with a guidance embedding set to about 3 at inference for single-pass generation. For training, guidance is set to 1, where CFG 1 is a no-op equal to CFG 0. Apply: Leave guidance at 1 during training, and use roughly 3 at inference to generate in one pass without classifier-free guidance.
Trigger-word style training for combined text-to-image and editing — Because Flux 2 does both text-to-image and editing, a style is taught on plain text-to-image data under a trigger word, so edit prompts such as 'make this picture [trigger] style' can invoke it. Apply: Train the style LoRA on captioned images with a trigger word, then test it with text-to-image prompts and with edit prompts that point to a reference picture.
Trigger-centred captioning with the [trigger] placeholder — Every caption starts with the trigger and describes only non-style content, such as 'a boat on the water with two men in it'. Style attributes like brush strokes, colors and paint go unmentioned, and anything uncaptioned binds to the trigger. Apply: Begin captions with [trigger], leave out clothing and hats if they should become part of the style, and describe them (for example 'dressed in 1980s style') if they should stay separate; the actual trigger can be chosen later.
Differential Output Preservation (DOP) — A preservation method in which captions containing the trigger are paired at runtime with a preservation-class version (for example 'Berthe Morisot style a man wearing a hat' vs 'photo a man wearing a hat'). With the LoRA dropped out, the base model's output on the partially noised image becomes the target the LoRA-on output must match. Apply: Set the trigger word and a preservation class (here 'photo') in the AI Toolkit job, so the model learns the style as the difference from its no-trigger output and keeps its existing knowledge elsewhere.
Blank Prompt Preservation — An alternative preservation method that runs a blank prompt and preserves what it produces. Unlike DOP, it can be combined with cached text embeddings. Apply: Use it instead of DOP when you want to cache text embeddings and unload the text encoder to save VRAM.
Text-embedding caching — AI Toolkit can cache text embeddings so the large text encoder can be unloaded during training. It cannot be enabled together with DOP, which alters every input and makes the tool throw an error. Apply: Cache embeddings only when not using DOP, for example when training a plain style without a trigger or when using blank prompt preservation, to cut VRAM use.
Float8 quantization of transformer and text encoder — AI Toolkit can quantize both the transformer and the text encoder to 8-bit. The speaker says this is required even on a 96 GB GPU, and training then needs about 64 GB. Apply: Set both the transformer and text encoder to float8 in the job configuration; the speaker notes that 48 GB still does not fit.
CPU offloading of text encoder or transformer — AI Toolkit can offload parts of the text encoder or transformer to CPU to lower VRAM, though the speaker does not know how low it can go. Apply: Enable offloading when VRAM is short, accepting that Flux 2 may still need a very large amount of system RAM.
Accuracy recovery adapters (planned) — A planned approach the speaker intends to try to see how low the VRAM requirement can go. Apply: The speaker plans to experiment with these to reduce the VRAM floor, which they compare to a limbo game.
LoRA rank selection for merged attention — The official Flux 2 implementation merges QKV into one linear layer, so most layers effectively need about 3× the usual rank, and rank 32 yields a LoRA that is not especially large. Apply: Set linear rank to 32 for this style LoRA.
Control images and the Match Target Res toggle — Control pairs are concatenated beside the target as in Kontext and Qwen Image Edit. Here controls resize down. With Match Target Res off, small controls arrive at native size, up to 1 megapixel combined. Apply: Add control pairs for edit training and leave Match Target Res at its default when not training controls; turn it off if you want controls at native size.
Mixed text-to-image and edit sample evaluation — Eight sample prompts: four plain text-to-image prompts to check how well the base ability is retained and four 'make this picture [trigger] style' edits of reference images (meme, cat, Cyclops Mona Lisa, dad photo). Apply: Include both prompt types in the sample set to see whether the style takes hold on edits without damaging text-to-image comprehension.
Dataset-bias mitigation — The speaker's suggested fixes when dataset skew changes subjects, such as turning a man into a woman: lower the LoRA weight, rebalance the dataset, or describe the subject in the prompt. Apply: Add more images of under-represented subjects, or prompt with details like 'a man with a mustache' when edits drift.
RunPod GPU setup for AI Toolkit — The speaker's cloud workflow: choose a large GPU (an RTX Pro 6000 with 96 GB, since 48 GB probably is not enough), switch the template to 'AI Toolkit official', set the AI Toolkit password variable, and set container disk to 300 GB (probably more than needed). The speaker also deploys on demand and opens the HTTP service link. Apply: Deploy the pod with those settings and open the HTTP service link to reach AI Toolkit.
AI Toolkit workflow and Hugging Face gated-model access — The in-tool sequence: create a dataset, drop in images and caption them (a blue caption is unsaved until you click it and press Enter), and add a Hugging Face read token in Settings after accepting the gated repo's terms. Then create a New Job with the model, quantization, rank, DOP and trigger set. Apply: Accept the Flux 2 terms on Hugging Face, add the token, build the captioned dataset, and queue the job; the model downloads automatically on first queue.
The speaker says a reference image in an edit model is processed side by side with the target. Four reference images roughly double the resolution and quadruple the pixels, which makes edit training memory-hungry.
Because the autoencoder has 32 channels, edits should keep unchanged pixels intact. The speaker says multi-pass encode/decode in other pipelines loses pixels even in regions that did not change.
The text encoder is a vision-language model (Mistral Small 3.1), not a pure language model like T5. The speaker argues this gives it better associations between concepts and visuals, which the speaker links to the model's understanding.
Quantized to float8, the 24B text encoder alone needs about 24 GB of VRAM to load. Because DOP prevents embedding caching, the encoder has to stay resident, which raises the VRAM floor.
Rank 32 gives a modest LoRA despite the 32B model. The official implementation merges QKV into one linear layer (and merges some other layers), so most layers effectively need about 3× the rank.
The speaker leaves clothing and hats out of captions on purpose so they bind to the style. Describing them in captions pushes them away from the style, though not perfectly.
Control images behave differently from Kontext and Qwen Image Edit. They are resized down rather than matched to the target's pixel count, so small controls (for example 256×256) can sit beside a larger target, up to 1 megapixel in total.
Dataset skew shows up in edits. The data was heavily white-women-heavy, so a male subject became a woman, and the dad sample lost his mustache. The speaker suggests lowering the LoRA weight, adding more men to the dataset, or describing the subject ('a man with a mustache') in the prompt.
Unrequested style features such as buns and flowers in the Cyclops Mona Lisa's hair became part of the style, which the speaker takes as a sign that the trigger word absorbs whatever the style data consistently shows.
In an earlier test on the developer demo model, training on the style alone did not preserve this kind of text-to-image ability. The speaker says the preservation setup is what made the difference.
The training run took 3,000 steps at an average of about 6.37 s/iteration (starting around 7–8 s). The speaker puts the total at a few hours to maybe half a day and says they have never trained an image model this large.
«A CFG of one is a non-op. Basically it's the exact same CFG as zero. That's how we train.»
— 01:34
«Not just a little big, it's very big.»
— 01:58
«I've said it before, it can't be done, and we did it anyway.»
— 03:40
«Anything not captioned the model's going to associate with the trigger word.»
— 14:11
«Even though this is an enormous model, rank 32 is actually not going to be that big of a LoRA just due to how the attention mechanism is merged.»
— 16:46
«Anytime this trigger word is not here, we want it to match what the model already knows.»
— 19:09
«I'm kind of learning this with you guys. And so, maybe we find out it doesn't work and you don't want to do it this way.»
— 22:53
«It's like the limbo game. We'll see how low we can go.»
— 24:48
«doing this differential output preservation really does help that you can really put all that information into that trigger word and separate it from everything else.»
— 31:27
«That's Flux 2. It's amazing. It's big. It's new. Zero-day release.»
— 32:39
Reception
Mostly supportive and engaged, with viewers thanking the creator and asking follow-up questions about Klein, learning rates and setup. A vocal minority criticized the video as unstructured and too talkative, and several people reported errors and access problems.
A release-day walkthrough of a Flux 2 style-LoRA run. Its most reusable content is the trigger-word captioning logic, the DOP-versus-embedding-caching tradeoff and the VRAM and quantization figures. The speaker presents the results as preliminary and notes dataset bias and open questions.

33:55