Generative art with language+diffusion models

3.1 Feature matrix

The table below is a quick-reference summary of the above; the licence and RAM columns are factual, the rest is aggregated community impression.

Speed is relative and hardware-dependent; published benchmarks usually run on a desktop NVIDIA GPU, which is faster than most laptop or Apple-Silicon setups. For live quality rankings — Elo scores from blind human votes, updated as new models ship — the Artificial Analysis text-to-image leaderboard is the canonical source.

3.2 Resolution

Think in a megapixel budget rather than fixed dimensions. Each model has a native total-pixel count it was trained around, and that budget is spent across whatever aspect ratio we choose — at ~1 MP, 1024×1024, 1216×832 and 1344×768 are the same budget in different shapes. This size-and-aspect conditioning is something SDXL introduced (Podell et al. 2023), training across multiple aspect-ratio buckets rather than a single square.

Two consequences follow. Generating much past native produces artefacts — duplicated heads and limbs, repeated motifs — rather than mere softness; the SDXL paper (Podell et al. 2023) motivates its whole design around the resolution limits of the 512-native Stable Diffusion 1.x models. And the route past native is upscaling, not a bigger generation: render at native, then enlarge with a tiled-diffusion pass or a dedicated ESRGAN-class upscaler (ChaiNNer and Spandrel are on the clients page).

How far each model holds together before it needs that upscale step varies by backbone:

• SD 1.5 is 512-native and falls apart not far above 768 — the limitation the SDXL paper (Podell et al. 2023) was written to address.
• SDXL and its fine-tunes (including Pony V6 XL) are 1024-native (Podell et al. 2023) (~1 MP), comfortable to roughly 1.5 MP.
• FLUX.1 dev / schnell sit at ~1 MP native and in practice stay coherent to around 2 MP (dev weights and card).
• FLUX.2 dev generates up to 4 MP natively — Black Forest Labs’ explicit pitch is detail at that size without an upscale step.
• Z-Image Turbo is trained at 1024 but generates to 2048×2048 (~4 MP) given the memory.
• Midjourney and Ideogram are fixed-output API services — preset sizes, not a megapixel dial.

A LoRA does not change any of this — it rides the base, so the niche Flux-dev LoRAs inherit Flux.1’s ~1–2 MP envelope regardless of what they were trained to draw.

On a memory-rich machine, the limiting factor is coherence, not memory: the question is what each model stays coherent at, and larger images cost time rather than memory. On a small machine, memory binds first, and output size competes with the model weights for it; the clients page covers the compromises.

3.3 Starting points by goal

• Photorealistic output for publication → FLUX.1 dev (or dev + LoRA)
• Fast iteration on 16 GB, no licence constraint → Z-Image Turbo, FLUX.1 schnell GGUF, or FLUX.2 [klein] 4B
• Specific aesthetic that probably exists as a fine-tune → check SDXL on CivitAI first
• Anime, creature design, dramatic poses, negative-affect content → Pony V6 XL
• Text legible inside the image → Ideogram (API) or FLUX.1 dev
• Just want something that looks good without thinking about it → Midjourney
• Need the deepest ControlNet or LoRA ecosystem → SD 1.5 (check SDXL first; if it’s not there, it may only be here)

With a model picked, the front-end clients page covers which app and runtime to run it through.

3.4 The rest of the menu

Open Draw Things and the model list is loooong — names like LTX, Wan 2.2, Hunyuan, ERNIE, HiDream-I1, Cosmos. Most of that length is the menu collapsing three different kinds of model into one alphabetical list. Video models (Wan 2.2, LTX-Video, Hunyuan) are a separate modality; for running them locally see the clients page. Cosmos is NVIDIA’s world foundation model for robotics and autonomous-vehicle simulation — it generates video, but as synthetic training data for embodied AI rather than as art, so for our purposes it is the one name to skip outright.

Among the image models proper, two more have ecosystems forming around them. HiDream-I1 (HiDream.ai, 17B, MIT) topped the Artificial Analysis board for a stretch of 2025 and has a modest but active LoRA following, with the lineage continuing in the O1 successor. ERNIE-Image (Baidu, 8B, Apache 2.0) is newer still — it occupies the text-rendering niche and picked up ComfyUI workflows and community quantizations within days of release, though it is too young to know whether the buy-in holds.

One prominent name is absent from the sketches despite filling a slot in every client’s menu: Stable Diffusion 3 / 3.5. SD3’s June 2024 launch went badly — weak anatomy, and a licence so restrictive that CivitAI temporarily banned all SD3 resources — and although Stability later revised the licence and shipped the improved 3.5, the community had already decamped to Flux. By the same community-action test, it has not earned a sketch: in 2026 its LoRA ecosystem is still thin beside SDXL or Flux. Being in the Draw Things menu and being something people actually build on are not the same thing.

4.1 Historical engravings and prints

• YFG Albrecht Dürer Engraving Style (Flux dev) — actually models Dürer’s burin line work (swelling/tapering strokes, cross-hatched form modelling) rather than the muddy “old print” pastiche most engraving LoRAs settle for. The closest thing to a proper copperplate idiom.
• WOOD ENGRAVING Style for FLUX (Flux dev) — clean large-block woodcut, not mezzotint sludge. Useful when we want Gustave Doré rather than Piranesi.
• gokaygokay/Flux-Engrave-LoRA (Flux dev) — gokaygokay is a well-known HuggingFace style-LoRA trainer with consistent technical quality. Good fallback if the CivitAI engraving LoRAs overfit on a particular subject matter.
• Perfect Ink Drawing — Engraving Style (Z-Image Turbo) — one for memory-constrained work: cross-hatched etching look on a fast base that runs comfortably in 16 GB.

4.2 Scientific diagrams

• YFG Patents (Flux dev) — trained on patent figures with numbered callouts, dashed hidden lines, exploded views. Most “blueprint” LoRAs only render whole machines; this one captures the schematic call-out idiom itself.
• Anatomica: Chalk (Flux dev) — anatomical chalk-plate aesthetic, and the creator notes the style transfers to non-anatomical subjects, so we get “Vesalius treats a toaster”.
• Century Botanical Illustration (Flux dev) — 18th-century plate aesthetic (stipple shading, latin labels, ivory ground). Almost every “botanical” LoRA returns modern watercolour; this one does engraved plates.
• Anatomical Surrealism (Flux dev) — Haeckel-meets-Da-Vinci hybrid of plate-style anatomy with mechanical/botanical.

Scientific-diagram coverage is thinner than one might hope. There is room here for a fine-tune on Ramón y Cajal-style neural plates or period microscopy.

4.3 Emoji and stickers

• fofr/sdxl-emoji (SDXL) — the canonical Apple-emoji LoRA. Lightweight, well-known, transparent-bg-friendly via post-matte.
• starsfriday/Kontext-Emoji-LoRA (FLUX.1-Kontext-dev) — different value proposition: style transfer to emoji. Feed a portrait, get the emoji version of that person. Kontext-native.

There is, as far as I can tell, no great isolated-icon-with-transparency model right now. Generate on a flat colour and matte out with rembg, BiRefNet, or SAM as a separate step.

4.4 Classic spot-colour posters

• Retro Ad Flux (Flux dev) — mid-century print advertisements, not movie posters: limited palette, halftone dots, Rockwell-ish illustration. The closest cousin to Cassandre and Saul Bass currently available.
• Doug Sneyd Illustration (Flux dev) — mid-century Playboy cartoon-illustration aesthetic. Fits the spot-colour-poster category sideways and has the confident flat-colour brushwork the era is known for.

No Saul Bass or Cassandre LoRA exists, for example. People approximate the look by stacking Retro Ad Flux with prompt directions (limited palette, cut-paper shapes, named reference).

4.5 Other specialist idioms

• Stained Glass Window Style (FLUX) (Flux dev, trigger SGW).

8.1 Alignment and the default outputs

Most production-facing image models are post-trained with RLHF (reinforcement learning from human feedback) or DPO (direct preference optimization) passes that align outputs with human preference ratings — Diffusion-DPO (Wallace et al. 2023) is a representative public example of the technique, and there is now a systematic survey of these methods for diffusion models. In practice, preference ratings correlate with positive affect and conventional aesthetics, and against content that tripped US-market human reviewers during training. The result is familiar to anyone who has tried to generate a mushroom cloud for an anti-war poster, a face that stays sad, or a creature design that reads as threatening. This is a training choice, not an architectural constraint. The same base weights with different post-training produce different behaviour; the open ecosystem exists, in part, because other choices are possible.

The client pipeline adds a second layer of conservatism — a CLIP-based classifier (CLIP being the image-text matching model that scores how well an image matches a text label) that fires after the image is generated and blacks out or blurs the result. These are separate constraints: a model with no alignment bias could still have a restrictive client wrapper, and vice versa. In diffusers-based pipelines the safety checker is a separately-loaded model component that can simply not be loaded; it is not part of the generation weights. The front-end clients page covers which clients load it by default; this section is about the models themselves.

8.2 Where the extra latitude comes from

Three separable things account for most of the gap between a cheery production endpoint and a model that renders what we asked. The capability and speed of the models named below are in the lineage sketches and feature matrix; what follows is only the latitude dimension.

Skipping the alignment pass. A fine-tune trained on raw character or creature art, without the RLHF passes of a production model, keeps a wider emotional and compositional range — negative affect, dramatic poses, non-cute creatures, the content a default checkpoint renders as cheerful regardless of the prompt. Pony Diffusion V6 XL is the canonical example. The range is a training difference, recoverable by anyone who post-trains differently — not a property of the architecture.

The separable safety checker. The output-side filter is client software, not model weights (the clients page covers which clients load it). On the model side, the base SD 1.5 / SDXL weights before any post-training carry fewer constraints than the fine-tunes every API provider ships — which is why providers ship the fine-tune, not the raw release.

Permissive licensing. Latitude in the licence is a different axis from latitude in the weights, and three regimes are in play:

• Apache 2.0 (FLUX.1 schnell, FLUX.2 [klein] 4B, Z-Image Turbo) places no restriction on subject matter at all.
• OpenRAIL — and its SD-family variant CreativeML OpenRAIL-M, which Stable Diffusion 1.5 used and Pony inherits — keeps the weights open but attaches a use-restriction clause prohibiting a listed set of harmful applications (child sexual abuse material, certain disinformation uses). It restricts what the model is used for, not who uses it.
• Non-commercial (FLUX dev, FLUX.2 [klein] 9B and [dev]) restricts commercial use of the weights, and Black Forest Labs ships a separate content-filter module alongside — often conflated with the weights, but distinct from them.

So the most permissively-licensed open weights are the Apache-2.0 ones, while the most permissive in content are the RLHF-skipped fine-tunes — and those need not be the same model.

CivitAI hosts the long tail of these fine-tunes (the Hugging Face vs CivitAI comparison covers the platform difference). Two alignment-relevant specifics: its content-gating is opt-in, so the default browse view is filtered but the full catalogue is not; and most uploads inherit their backbone’s licence, so an SD-family fine-tune is usually CreativeML OpenRAIL-M whether or not its page says so.

8.3 Negative-valence imagery

RLHF alignment correlates with positive affect on underspecified prompts — this has theoretical backing in the preference alignment literature (Wallace et al. 2023) but AFAICT has not been studied specifically for image emotional valence; treat it as well-grounded scuttlebutt rather than settled science. Community prompt engineering guides report that explicit physical descriptors work better than emotional labels: “tears streaming, jaw clenched, hollow eyes” rather than “sad character.” Adding positive-affect terms to the negative prompt — happy, smiling, cheerful, resolved, peaceful — is a commonly-reported workaround; the mechanism would be steering the sampler away from the part of latent space reinforced by positive preference ratings. I have found this useful but have no formal evidence it generalises.

The behaviour varies by model family and post-training regime. Fine-tunes that skipped RLHF passes respond differently; Pony Diffusion is the main example with a substantial community behind it.

For politically or historically charged imagery — mushroom clouds, battle scenes, burned buildings — there is a further distinction. Client-level classifiers (the safety_checker, InvokeAI’s NSFW blur) fire on the output image, not the prompt; the model weights themselves often do not have the corresponding restriction. Framing such prompts in historical or clinical registers (“nuclear test, Bikini Atoll, 1952 archival photograph, monochrome, fallout plume”) is anecdotal community knowledge — I have seen it reported to help bypass output classifiers, but have no systematic evidence for it, and how durable it is against classifier updates is unknown.

8.4 Jurisdiction

The legal landscape for training on copyrighted images, running inference, and publishing outputs differs across the jurisdictions relevant to a researcher working between Asia, Australia, and Europe. What follows is a snapshot as of mid-2026; case law is thin in all four regimes and several questions are actively contested. One term recurs throughout: a TDM (text and data mining) exception is the copyright carve-out that lets a model train on in-copyright works without first licensing them.

Japan: The interpretive question is whether inference “for the enjoyment of the expression” falls outside Art. 30-4’s scope. The 2024 Agency for Cultural Affairs guidance is advisory, not binding case law. Publisher suits against AI companies (Asahi/Mainichi v. Perplexity, filed September 2025) may produce clearer precedent; outcomes are pending as of this writing.

Singapore: Section 244 was modelled on Japan’s Art. 30-4 and is explicitly non-waivable — copyright owners cannot eliminate it via their terms of service (ToS), which matters when using images from platforms with “no AI training” clauses. The May 2026 Online Criminal Harms Act guidelines target AI-generated NCII of real persons specifically, not artistic or scientific generation.

EU: C2PA metadata (Content Authenticity Initiative) is the emerging standard for Art. 50 compliance — machine-readable provenance embedded at generation time. The c2patool CLI is the open-source implementation. Non-compliance from 2 August 2026 attracts fines up to €35M or 7% of global annual revenue.

Australia: The most restrictive of these four jurisdictions for training — no TDM exception and no general fair use doctrine, only enumerated fair dealing categories. For published research outputs, the deepfake provisions in the Criminal Code and state laws are expanding rapidly (three state-level laws passed between November 2025 and early 2026); the common thread across all of them is that they target distribution of intimate imagery involving real persons, not research generation. The s.41A parody exception exists and has been tested (Universal Music v Palmer), but courts will want evidence of human satirical intent, which creates an awkward question for AI-generated outputs where the human’s role was primarily prompt engineering.

8.5 Hygiene

Several things seem to reduce exposure, per practitioners and lawyers in this area; none is a guarantee.

Documenting purpose before generating — a paragraph in a lab notebook or file header — creates contemporaneous evidence of intent. Fair-use and fair-dealing analysis in all four jurisdictions makes purpose and intent explicit factors.

Keeping prompt logs for published work has become a practical precaution. Courts have begun ordering preservation of AI generation records specifically: the SDNY’s May 2025 order in New York Times v. OpenAI required OpenAI to preserve “all output log data on a going forward basis”; a National Law Review analysis (Feb 2026) notes that GenAI prompts, outputs, and logs are now treated as discoverable ESI (electronically stored information) under standard rules. Retrospective reconstruction is harder to defend than contemporaneous records.

Embedding C2PA provenance metadata at generation time — mandatory in the EU from August 2026 — is useful elsewhere for the same reason: it makes the origin of an image legible without relying on the viewer to trust a caption. Draw Things and some ComfyUI node packs support it; the c2patool CLI works on any file.

Distributing NSFW or dramatically charged research imagery only within a research team or behind access controls is a common precaution. All four jurisdictions hinge liability primarily on distribution, not generation, which means the same image can be lower-risk as an unpublished research artefact than as a blog illustration.

Identifiable real persons in embarrassing or intimate contexts carry civil and — in some jurisdictions — criminal exposure regardless of artistic intent.

If our work may reach multiple jurisdictions, using EU Art. 50 requirements as a floor (watermark + human-readable “AI-generated” disclosure) satisfies all four regimes simultaneously.