Creative AI Briefing September 17, 2026 morning edition 13 min read

Nothing new arrived in the models today, but three separate groups deleted a required piece of the machinery around them, and every deletion lands on your card as free memory or on your clock as saved time.

JiT-DDT Linum Ideogram v4 Krea 2 MiniMax H3 FastVideo FastH3 ClipProj image-gen video-gen local-creative-ai open-weights prompt-craft licensing-provenance comfyui

Creative AI Briefing: Thursday, September 17, 2026

Type a caption as a structured design brief instead of a sentence, hand it to a 4-bit copy of Ideogram v4 on a 24 GB card, and get a typographic poster back in eight steps. That went up yesterday morning. Look at everything else that shipped in the last two days and the same shape keeps repeating: nobody released a new model, and three separate groups deleted a required component out of an existing one. An image model without the compression stage. A video model without its 15.7 GB text encoder. A design model without three quarters of its file size. The models are not getting better this week. The scaffolding bolted to them is getting thrown away, and every piece thrown away lands on your card as free memory or on your clock as saved time.

New models

JiT-DDT from Linum is a 2.5-billion-parameter text-to-image model with no VAE in it at all (weights, Hugging Face createdAt September 14, 21:25 UTC; write-up published September 16). Nearly every image and video model you use, FLUX, Krea 2, Z-Image, splits the job in two: a VAE squeezes pixels into a compact code, and a second model learns to generate inside that code. Linum threw the first half away and made the generator do its own compressing. The claim, measured against their own previous model, is 3.6 times fewer GPU-hours and 4.2 times fewer training samples, at four times the pixels: 512 × 512 out, against 256 × 256 before.

The architecture is two halves that specialise. An encoder half predicts a 64 × 64 version of the picture, which is to say the structure with all the detail stripped out. A decoder half takes that sketch, its own view of the noisy input, and the text prompt, and paints the full 512 × 512. That division exists because removing the compression stage made detail worse, not better. Linum say so plainly: models built this way are known to struggle with fine detail, their own tests confirmed it, and their first attempt produced airbrushed faces and oversaturated colour. The two-half split, plus a change to the training noise schedule and some extra blocks that let picture and text settle before mixing, is what clawed the detail back. Worth holding onto, because the headline version of this release will be "they deleted the VAE and it got better," and that is not what the write-up says.

Linum are blunt about what this is. The card calls it "a research artifact, not a full model release," trained on 138 million samples with no post-training. It is a 9.95 GB fp32 file, samples at 50 Euler steps, and needs Qwen3.5-4B alongside it to read your prompt. Code and weights are Apache 2.0, which is as permissive as it gets. There is no hosted demo, and the GitHub repo sits at zero stars, so nobody outside Linum has kicked the tyres yet.

Image

Ideogram v4 now fits on a 24 GB card. Four quantized builds went up yesterday morning from rootonchair: ideogram-v4-instant-nunchaku-lite-int4 and its NVFP4 twin (both createdAt September 16, 08:58 and 08:59 UTC), plus the same pair for the Fast variant (08:54 and 08:57). The int4 build is roughly 16 GB against a bf16 original of 9.28 billion parameters. Ideogram is the model people reach for when a picture has to carry readable type, so getting it onto a consumer card matters for anyone making posters, packaging comps or title cards.

Two catches, both stated plainly on the card. The first is a licence: the upstream weights at fal/ideogram-v4-instant are gated behind the Ideogram Non-Commercial Model Agreement, so this is for learning and testing, not for client work. The second is a prompting requirement almost nobody covers: Ideogram v4 expects a structured JSON caption, not an English sentence. Ideogram's own prompting guide says the model was trained exclusively on these captions and that plain text "will not work and will likely trigger a safety warning." You get bounding boxes on a 0 to 1000 grid, up to 16 hex colours for the image and 5 more per element, and every piece of in-image type declared as its own object. It reads like a layout spec because it is one. See the workflow below.

Krea 2 gets pixel art at three resolutions. e-n-v-y/Krea-2-Pixel-Art (createdAt September 16, 15:06 UTC, MIT) ships 32 × 32, 64 × 64 and 128 × 128 add-ons, generated at 1024 × 1024 and cleaned up by a companion node, ComfyUI-Krea2-Pixel-Art-Refiner. Trigger word is pixel art, except on the 64 × 64 file, where the author says it works better without it. MIT on the add-on; Krea's own terms still govern the base model.

Video

No lab shipped anything. What moved was the local tier of MiniMax H3, and it moved twice in one day.

Someone put the 8-step FastVideo build on Apple Silicon: vanch007/FastVideo-FastH3-8-Step-V2-MLX-INT8 (createdAt September 17, 07:49 UTC, Apache 2.0), with a 6-bit sibling forty minutes earlier. It is about 22 GB down from 65 GB, and the author reports roughly 31 minutes end to end on an M3 Max with 128 GB for 124 frames at 480 × 832 and eight steps. Recommended hardware is an M2/M3/M4 Max or Ultra with 64 GB or more. Half an hour for five seconds of video is not interactive, but it is a Mac with no cloud bill.

Separately, realrebelai/FastH3-V2_GGUFs (createdAt September 16, 03:27 UTC) packaged the same base for ComfyUI on smaller NVIDIA cards.

Open and local

Today's local story is subtraction. Every item here makes an existing model need less, rather than do more, and most of them publish the numbers behind the claim.

Linum-AI/jit-ddt: a text-to-image model that does its own compression, so there is no separate VAE file to load or fine-tune around. Why now: createdAt September 14, Apache 2.0, with a long public write-up dated September 16. The GitHub repo at Linum-AI/jit-ddt sits at 0 stars, so this is very fresh and untested by anyone else.
rootonchair/ideogram-v4-instant-nunchaku-lite-int4: typography-grade image generation at roughly 16 GB. Why now: createdAt September 16, with bf16 / int4 / NVFP4 sample images published side by side and the author noting small text is slightly worse than the original.
vanch007/FastVideo-FastH3-8-Step-V2-MLX-INT8: MiniMax H3 video on a Mac, 22 GB instead of 65. Why now: createdAt September 17, Apache 2.0, with the measured 31-minute figure and the exact command published.
e-n-v-y/Krea-2-Pixel-Art: game-ready sprites and tiles from Krea 2 at three pixel grids. Why now: createdAt September 16, MIT, and it ships with the post-processing node rather than leaving you to clean up the grid yourself.
SiCl4kksk/anima-turbo-4step: a four-step ONNX build of Anima, for people who want illustration output without a Python stack. Why now: createdAt September 16, 08:14 UTC.
jhong520/ClipProj-MiniMax-H3: a same-day mirror (createdAt September 16, 03:27 UTC) of work that is not new. The original is NicoLab28/ClipProj-MiniMax-H3, createdAt August 9, 154 likes. Worth naming anyway because of what it does, below.

Creative workflows

1. Write an Ideogram brief instead of an Ideogram prompt.

This is the one that will change your output today, and it costs nothing to try. Ideogram v4 does not want a sentence. It wants a small structured document, and the model card spells out the shape.

The steps. The full schema is published at ideogram-oss/ideogram4/docs/prompting.md. Three top-level fields, only the third required: high_level_description, a sentence or two; style_description, holding aesthetics, lighting, medium, either a photo or an art_style field, and an optional palette of up to 16 uppercase hex colours; and compositional_deconstruction, holding a background string and an elements list. Each element is obj or text, takes an optional bbox in [y_min, x_min, y_max, x_max] on a 0 to 1000 grid, a desc, and up to five hex colours of its own. A text element also carries the literal string to render, in its own field. Key order inside each object is strict and the pipeline's caption verifier warns when you get it wrong. If you do not want to hand-write this, the repo ships a free magic-prompt path that expands a plain sentence into a valid caption server-side.

How it works. The guide says it outright: the model was trained exclusively on these structured captions, and "passing in plain-text prompts directly to the model will not work and will likely trigger a safety warning." The bounding boxes are the part worth sitting with. You are not describing a picture and hoping. You are placing a headline in the top third and a logo in the lower left, on a coordinate grid, the way you would in any layout tool.

Why it is good. It rewards what a designer already does: separate content from treatment, list the pieces, place them. It re-rolls cleanly, because changing one element's description leaves the rest of the composition mostly intact. And the palette control is real typographic control, not a mood word.

Where it breaks. Small text is still imperfect, and the 4-bit build is slightly worse at it than the original, so set type at poster scale. Hex must be uppercase full-length #RRGGBB; shorthand is rejected. The safety filter returns a grey "Image blocked by safety filter" card, and Ideogram say the false-positive rate is high specifically for non-JSON prompts, which is one more reason to write the caption properly. The local weights are non-commercial, so a paying job goes through Ideogram's own product.

2. Get 11 GB back on your card by swapping out a video model's text encoder.

This is not new work, but it is the most useful thing in the local tier right now and it went into circulation again yesterday via a mirror, so it is worth stating clearly. Full credit to NicoLab28, first published August 9.

MiniMax H3 loads a 32-billion-parameter Qwen3-VL just to turn your prompt into numbers: 15.7 GB of your card doing nothing but reading text. ClipProj is a set of small translation files that let a 4B or 8B Qwen3-VL do that job instead. Reported cost on the card: 15.7 GB down to 4.5 GB, with no change to the video model, the VAEs or the sampler.

The steps. Install the node from github.com/nicolab28/ComfyUI-ClipProj (151 stars). Put the projection files in ComfyUI/models/clip_projections/. Start with mmh3-4b-ClipProj-v3.1, 26 MB of projection plus about 4.6 GB for its encoder. Before you trust anything, run the two control files, mmh3-ClipProj-control-zero and mmh3-ClipProj-control-identity, on the prompt "a red ball on a wood table." The zero control should return a countryside landscape and the identity control an unusable golden object in flames. If either looks fine, the real file is not doing the work and you want to know that first.

How it works. Every Qwen3-VL shares one tokenizer, so a prompt breaks into the same pieces at the same positions in a 4B and a 32B. That makes it possible to fit a plain position-by-position translation between what the small model produces and what the big one would have produced. The main file is fitted by ordinary regression, no training loop at all.

Why it is good. It is 11 GB of headroom on the exact card you already own, it is MIT, and the author published the failures alongside the wins, including three of his own claims he later corrected. The best of those corrections is a lesson for anyone who reads model cards: a change that moved his similarity score from 0.697 to 0.845 turned out, on inspection, to produce a mathematically identical file. The whole gain was an artifact of measuring in a different space, and he wrote that up rather than deleting it and saying nothing.

Where it breaks. Named people were the original weakness and the -celeb files fix most of it, but characters identified by a mask or a costume still come out as a stranger in the right outfit, and people famous before photography come out generic. The measured similarity to the original conditioning tops out around 0.80, which sounds bad and mostly is not, but the author is clear that the number does not predict which file renders better, so test both the plain and the residual version on your own prompts. And the legal status of the files themselves is, in his own words, unclear.

Worth testing

Ideogram v4 Instant on Hugging Face, the hosted Space linked from fal's repo, if you want to test the JSON-brief technique without downloading 16 GB. Tradeoff: it is someone else's Space, so availability is not guaranteed, and the underlying licence is still non-commercial.
Krea 2 pixel art via the MIT add-ons. Tradeoff: no hosted demo, and you need the refiner node or your grid will be mush.
JiT-DDT if you have a card and curiosity. Tradeoff: 512 × 512 only, no post-training, no demo, and zero stars on the repo means you are the test.
ClipProj if you already run MiniMax H3 locally. Tradeoff: run the controls first, and expect a bad day with named characters.
FastH3 on MLX if you are on a 64 GB-plus Mac. Tradeoff: about 31 minutes per clip on an M3 Max, so this is an overnight-render tool, not an iteration tool.

What actually matters from today's signal

The labs published nothing creative in 48 hours, and that absence is doing something useful: it makes visible how much of the cost of generating anything is not the model. A VAE, a text encoder, a set of full-precision weights, a Python stack. Today three different people, none of them at a frontier lab, removed one of those from a working pipeline and published the receipts. The pattern is not compression for its own sake. It is that the required parts are turning out not to be required.

The practical consequence is that your next capability upgrade may not be a download at all. It may be reclaiming 11 GB you were spending on a text encoder, or discovering that the model you already run has a prompt format you were never using. The Ideogram JSON brief has been the documented interface the whole time, and I would bet most people generating with it are still typing sentences and blaming the model for the typography.

The counter-signal is worth naming. Two of the three best things here are legally awkward. The Ideogram weights are gated and non-commercial upstream and openly mirrored downstream. The ClipProj author says outright that the legal status of his files is unclear. Only JiT-DDT and the Krea add-on come with a licence you could show a client without flinching, and JiT-DDT is a half-finished research preview by its own admission. So weigh these by what you are actually going to do with them. For learning how a model works, all of it is free. For invoicing, the list gets very short very fast.

Source access notes: openai.com/news, blog.adobe.com, bfl.ai/blog, blog.fal.ai, replicate.com/blog, stability.ai/news, runwayml.com/research and elevenlabs.io/blog all fetched successfully and had nothing creative inside the 48-hour window; the most recent creative items are Adobe's Premiere and After Effects AI update dated September 8, Runway's GWM Worlds 2 on September 3, fal's H3 Max on August 27 and BFL's FLUX Upscale on August 20. Stability's news index renders without dates and contributed nothing datable. blog.comfy.org was not reached; the ComfyUI GitHub releases API returned an empty body again this run and no ComfyUI release is claimed here. Midjourney and Civitai were not reachable and are not cited. The sandbox has no outbound network access for curl, so every fetch went through the fetch tool; Hugging Face API calls were cache-busted. Two date traps were caught and excluded: stabilityai/stable-audio-3-optimized showed as updated within 24 hours on the listing page but its API record gives createdAt May 18 and lastModified September 1, and KlingTeam/ShotStream appeared 16 hours old on the listing but its own repository log dates the release to March 27, 2026. Both are card refreshes, not releases. Every Hugging Face date above comes from the API createdAt field; star counts come from shields.io with cache busting. An adversarial fact-check subagent reviewed this draft against primary sources and caught two errors, both corrected above: a self-correction anecdote about a 0.15 similarity jump had been attributed to Linum when it belongs to the ClipProj author, and the Ideogram agreement was named with a "4" its own title does not carry. The subagent did not independently re-confirm the Runway, fal and BFL dates in this note.