Creative AI Briefing: Wednesday, September 23, 2026

Draw a rough scribble of a figure walking across a room, render it out as a short black-and-white clip, and hand that to a frontier video model as the instruction. It follows your drawing. That shipped Tuesday. Look at the other four things worth your attention from the same 48 hours and the pattern is uncomfortable for anyone who reviews models on launch day: not one of them is a model. A control branch for MiniMax H3. A weight merge that fixes MiniMax H3's quality tradeoff. A patch adapter for a Qwen image model that has been out for a week. A pair of style dials, one of them for an image model that landed on Monday. The labs supply the raw material now, and the layer that makes it directable arrives from strangers a few days later.

New models

MiniMax-H3-Fun-Controlnet-Union-2.0 went up Tuesday, September 22, from Alibaba's PAI group, and it is the single most useful thing in this briefing. It is not a video model. It is a control branch that bolts onto MiniMax H3, the frontier open-weights video model, and lets you condition the output on a control video instead of only a prompt. Version 1 arrived August 24 with five control types. This one has eight, and the three new ones are the interesting ones: Scribble (hand-drawn lines), Layout (color-coded bounding boxes on white, so you place each subject in frame by drawing a rectangle), and Gray (a luminance pass, so you keep the exposure and contrast of a source clip while replacing everything in it). Canny, Depth, HED, MLSD and Pose carry over. One checkpoint covers all eight, so no swapping files between a pose shot and a depth shot.

The mechanism that matters: the control branch now injects into ten of the base model's fifty transformer blocks rather than five, roughly double the attachment points, which the model card frames as tighter structural adherence. A control_context_scale dial runs from 1.0 (follow the control video hard) down to 0.0 (ignore it). Video inpainting also changed recipe: masked pixels now sit at mid-gray after normalization rather than near-black before it, and the author says filled regions blend into kept regions better as a result.

Three honest catches. The file is 13.5 GB of control weights on top of the base model, and the card states that the transformer at roughly 62 GB plus the Qwen3-VL prompt reader at roughly 62 GB will not fit an 80 GB card fully loaded, so you run it with offloading. The checkpoint is guidance-distilled: set guidance above 1.0 and you apply guidance twice and degrade the output. And loading the version 1 config against this file is documented as a silent failure that drops half the control weights and gives you wrong output with no error. Licence is the MiniMax H3 Community License Agreement, with territorial restrictions and an acceptable-use policy you should read before commercial work. Code lives in VideoX-Fun. No hosted demo exists for the control branch; the base model is playable at fal and Krea if you want to judge the underlying image quality first.

Video

Cobijada MiniMax-H3 Hybrid landed Tuesday evening and solves an annoying forced choice. MiniMax H3 ships in two flavours with identical architecture: FL, keyed off a first and last frame, which looks better; and REF, which accepts image, video and audio references, and which the author says costs real quality even on jobs that use no reference at all. So you pick between reference control and output quality.

Winnougan compared the two checkpoints tensor by tensor and found they are nearly the same file. The meaningful difference sits in the per-block conditioning projections that route the conditioning signal into each layer. So the merge takes FL as the base everywhere and swaps in REF's conditioning projections for blocks 25 through 49 only. No training, no fine-tune, just weight selection. The claim is that you keep REF's reference pathway while holding FL's fidelity elsewhere, and the author is careful to say it will not beat FL on non-reference work. Ships in full precision plus two smaller quantized variants, drops into ComfyUI/models/diffusion_models/ with the stock loader, no custom node. The merge and the conversions are Apache 2.0; the underlying H3 weights keep MiniMax's own terms.

Image

Qwen-Image-2.1-Fix is a 28-like adapter posted Tuesday that claims to fix "most of the image generation issues" with Qwen-Image-2.1, which went up on September 14. The card is thin and the claim is the author's own, backed by fifteen side-by-side comparison renders in the repository and a workflow zip with the settings the fix depends on. Treat it as a community report rather than a measured result, and judge it on the comparison images. What it signals is more interesting than the file: a frontier image model picked up a third-party corrective patch, with a following, inside eight days of release.

ntc-ai's particle sliders arrived early Wednesday for Supra2-IMG, the 417 MB image model that appeared Monday, and a few minutes later for Krea 2. These are style dials rather than style presets. Turn "Final Boss" up and your plain service robot grows angular shoulder armour and weathered panels in the same yard, same prompt, same seed. The Krea 2 files are ordinary adapters that load with stock Load LoRA at model strength 1 and text strength 0, about 77 MB each, with half-size compressed versions at roughly 38 MB. The Krea files carry the Krea 2 Community License, so check that before selling anything made with them; the Supra pair is Apache 2.0.

The Supra release also publishes a number worth sitting with: 1,600 training updates in 134.4 seconds on one rented A6000, with 1,541 MiB of peak allocated memory, producing a 6.8 MB dial. The author is scrupulous about what that does and does not mean, noting the converged version took about 47.5 minutes and that the two-minute figure describes a single timed run, not convergence. Still, the floor for making a custom style control on a small model is now measured in minutes.

Open and local

The local story today is that pruning and merging have become ordinary craft rather than a specialist act. Two of the five items above are weight surgery performed by individuals on models they did not train, published with the method written out. Neither required a training run.

• aigc-apps/VideoX-Fun: the inference code for the new eight-condition control branch. Run examples/minimax_h3_fun/predict_v2v_control.py for control, predict_v2v_control_inpaint.py for masked regeneration (repo). Why now: version 2.0 of the control weights dropped Tuesday and the config file requirement changed with it.
• HyperGAN/supra-particle-sliders: training and inference scripts for concept dials on the tiny Supra image model, with timing evidence as JSON (repo). Why now: published Wednesday alongside the weights.
• HyperGAN/krea2-particle-sliders: the same recipe against Krea 2 Turbo, with a ComfyUI node and setup guide (repo). Why now: the Krea sliders posted a few minutes after the Supra pair.
• ali-vilab/VACE: not new, but it is the preprocessing dependency for the new Layout condition. Its vace_preproccess.py --task layout_track turns a reference video or a pair of boxes into the coloured-box control video the new checkpoint expects (repo).

Star counts are omitted here because the shields.io endpoint was not fetched this run; treat these as verified repositories, not as ranked projects.

Creative workflows

1. Block a shot with rectangles instead of adjectives. Direct link: the Layout condition section of the 2.0 model card, driving examples/minimax_h3_fun/predict_v2v_control.py in VideoX-Fun.

The steps. Produce a layout video: per-subject bounding boxes rendered as colour-coded rectangles on a white background, one colour per subject, moving across the frames the way you want each subject to move. You can track them from a reference clip with VACE's vace_preproccess.py --task layout_track, or you can hand it a pair of boxes directly. Point control_video at that file, set config_path to config/minimax_h3/minimax_h3_control_inpaint_post_norm.yaml, write a prompt describing scene, subject and camera, and run.

How it works. The layout video is an ordinary RGB video as far as the model is concerned. The control branch reads it at ten points through the network and pulls the generation toward that structure. Because a rectangle carries position and size and nothing else, you are specifying staging while leaving everything else to the prompt.

Why it is good. Staging is the thing prompts are worst at. "On the left" is a coin flip; a box at x=230 is not. This also gives you something to show a client before any pixels exist.

Where it breaks. Boxes carry no pose, so a subject can be in the right place doing the wrong thing. The output resolution and aspect ratio are taken from your control video, and frame count snaps down to the largest value the decoder accepts, capped at 15 seconds at 24 fps. And the config trap above is real: the wrong YAML gives you a plausible-looking wrong render, not an error.

2. Regenerate one region of a shot and keep the rest. Direct link: the inpainting section and examples/minimax_h3_fun/predict_v2v_control_inpaint.py.

The steps. Make a mask video the same length as your source: white where you want new content, black where the original must survive. Feed source and mask to the inpaint script with the post_norm config, prompt the replacement, render.

How it works. The control input carries three things at once: the encoded source, the encoded source with the masked region removed, and the mask itself. Version 2.0 changed where the hole is zeroed, so the empty region now sits at mid-gray instead of near-black, which the author reports makes filled areas blend into kept areas more cleanly.

Why it is good. This is the video equivalent of a paint-out, and it does not require you to re-render the whole shot to change a sign, a garment or a piece of set dressing.

Where it breaks. Masks must be authored by hand or by a separate segmentation pass, and the model card publishes no numbers on edge fidelity. A soft or shaky mask edge will show. Nothing here preserves audio.

3. Train a style dial in an afternoon on a laptop-class budget. Direct link: the training formulation and reproduction notes.

The steps. Write six matched prompt pairs, neutral against the style you want. Supply no images at all. The script renders from both sides of each pair and teaches a small adapter to move the neutral render toward the styled one, mixing in a preservation prompt (a lake, a bicycle, a cat) every fifth update so the dial does not bleed into unrelated subjects.

How it works. The style is defined by the difference between two prompts rather than by a folder of reference pictures, which is why no dataset is needed.

Why it is good. No image collection, no licensing question about training material, no rights clearance. The output is a small file you can turn up and down.

Where it breaks. The author says plainly that the dial can shift unrelated objects, that the compressed half-size version weakens the effect, and that the fast timing was one run on one card and is not a benchmark. Supra is a 256-pixel model, so faces, fine detail and lettering are poor, and stray text appears in dark scenes.

Worth testing

• The Gray condition on the new control branch. Take a graded clip you like, feed its luminance as the control, and generate entirely new content that inherits your exposure and contrast. Tradeoff: you get no hosted demo, so testing costs a rented 80 GB card and an offloading config.
• Cobijada's hybrid merge as a drop-in for the REF checkpoint. If you already run reference-conditioned H3 workflows in ComfyUI, this is a file swap. Tradeoff: no published comparison renders, so the quality claim is untested by anyone but the author.
• Krea 2 particle sliders through stock Load LoRA. Cheapest item here to evaluate: about 77 MB, no custom node, and the card shows off-versus-on pairs at fixed seeds. Tradeoff: the Krea 2 Community License governs the weights, so read it before commercial work.
• The Qwen-Image-2.1-Fix adapter and its workflow zip. Tradeoff: the settings matter as much as the file, and the improvement claim is self-reported with no measurement behind it.

What actually matters from today's signal

The useful unit of creative AI is no longer the model. It is the attachment. Three of today's five items are files that do nothing on their own and everything when loaded on top of something else, and the two that are not are weight merges of someone else's checkpoints. A working creator should stop reading launch posts as though they were product reviews. A model on launch day is a slab of capability with a prompt box in front of it. The control surface that makes it usable for actual shot work arrives two to four days later, from people whose names are not on the announcement.

Watch the money side of that. The attachment layer is where cost concentrates now. The control branch is 13.5 GB you have to load alongside 124 GB of base model and prompt reader, which is a rented-card job no matter how cheap the adapter was to make. Meanwhile a style dial costs two minutes and produces a 6.8 MB file. Image work has become nearly free to customize; video work has become directable and stayed expensive. If you plan a project this quarter, that asymmetry is the planning constraint.

The counter-signal is licensing, and it is getting worse rather than better. Today's most capable item carries a community licence with territorial restrictions. The Krea sliders carry another. The Cobijada merge is Apache 2.0 for the merge itself and something else entirely for the weights inside it. That stacking is now normal: three licences deep for one render, each one added by a different party, none of them written by the person who will be asked whether the output can ship. Nobody is going to untangle that for you. Read the licence on the base model before you build a workflow on top of an adapter, because the adapter's permissive tag does not travel down.

---

Source access notes: An adversarial fact-check pass ran against primary sources before publication. It caught four errors, all fixed here: Qwen-Image-2.1 was dated from a listing modification timestamp rather than its API createdAt (real date September 14, not September 20); the lead item's ship day was stated two ways in one document (Tuesday is correct); the Supra base model was described as a Sunday release when its createdAt says Monday; and a peak-memory figure was rounded down rather than quoted. It confirmed the file sizes, block counts, condition counts, licence names, the August 24 date for version 1, and the timing figures. Blocked or empty sources this run: blog.comfy.org returned a JavaScript wall with no post content; the ComfyUI GitHub releases API and the repository CHANGELOG both returned empty bodies, so no ComfyUI core release is reported today. openai.com/news and blog.fal.ai were reachable and carried nothing creative-facing in the window (OpenAI's Sep 22 items are text-model and caching posts). Replicate and ElevenLabs blogs were reachable with no posts in the window. Adobe, Runway, DeepMind and Midjourney were checked via WebSearch with no releases surfacing in the window; midjourney.com/updates was not fetched directly. Star counts were not fetched from shields.io this run and are therefore omitted rather than estimated. Every Hugging Face date here comes from the API createdAt field, not a listing timestamp. No hosted demo exists for the lead item; the substitutes named are for the base model only. The Qwen-Image-2.1-Fix improvement claim and the Cobijada quality claim are both self-reported by their authors and labelled as such inline.