Stable Diffusion 3.5

Overview

Stable Diffusion 3.5 is Stability AI's flagship open-weights image model family, announced on October 22, 2024. It is built on the Multimodal Diffusion Transformer (MMDiT) architecture — a departure from the U-Net backbones of SD 1.5 and SDXL — and ships in three variants tuned to different quality, speed, and hardware budgets.

Nearly two years on, it remains current. As of September 6, 2026 Stability's Stable Image page still describes Stable Diffusion 3.5 as "our most powerful image model yet," and no successor image model has been announced or released. There is no Stable Diffusion 4; pages claiming an SD4 launch cite no Stability announcement.

That is not the same as a dormant company: Stability's 2026 output went into other product lines — Brand Studio, Stable Audio 3.0, an August funding round backed by the major music labels — covered under Ecosystem & Tools. It does mean the image line has had no architectural successor for close to two years, and the benchmarks below show what that costs.

The weights are released under the Stability AI Community License — free for non-commercial use, and free for commercial use by organizations under $1M in annual revenue.

Variants

Note that Medium generates at a higher maximum resolution than Large (2 MP versus 1 MP), while Large leads on prompt adherence. The variants trade along different axes rather than forming a straight quality ladder. Stability describes Medium as outperforming other medium-sized models on the balance of prompt adherence and image quality.

Capabilities

Stability's three model cards open with the same sentence, changing only the variant name: each is an MMDiT text-to-image model "that features improved performance in image quality, typography, complex prompt understanding, and resource-efficiency." What differs is how each variant gets there.

• Prompt adherence: Stability's headline claim for SD 3.5 Large is that it "leads the market in prompt adherence and rivals much larger models in image quality" — the ability to render every element of a complex prompt, including multiple subjects and their spatial relationships. No number is attached to that claim anywhere in Stability's documentation; see Performance.
• Typography: Named on every model card as one of the four things SD 3.5 improved. Stability publishes no text-rendering accuracy figure.
• Three fixed text encoders: OpenCLIP-ViT/G and CLIP-ViT/L, both at a 77-token context, plus T5-XXL at "77/256 tokens at different stages of training." The Medium card warns that "you may observe artifacts on the edge of generations when T5 tokens go over 256."
• Four-step generation (Large Turbo): Adversarial Diffusion Distillation, which the card says "allows sampling with 4 steps at high image quality." The reference diffusers call uses num_inference_steps=4 and guidance_scale=0.0 — no classifier-free guidance at all.
• Multi-resolution generation (Medium): MMDiT-X "introduces self-attention modules in the first 13 layers of the transformer, enhancing multi-resolution generation and overall image coherence," trained through progressive stages of "256 → 512 → 768 → 1024 → 1440 resolution." Stability recommends "sampling with Skip Layer Guidance for better structure and anatomy coherency." The reference call is 40 steps at guidance 4.5. (The same card describes the change as "dual attention blocks in the first 12 transformer layers" two paragraphs earlier; Stability's own count is inconsistent.)
• Consumer-hardware generation: Medium is explicitly targeted at consumer GPUs, at 9.9 GB VRAM excluding text encoders.
• Stylistic range: The family covers photorealistic, cinematic, illustrative, and stylized output, and Stability emphasizes diversity of representation in generated outputs.
• Fine-tuning friendly: Query-Key Normalization in the transformer blocks, "stabilizing the model training process and simplifying further fine-tuning."

One boundary is stated on the model cards rather than inferred: "The model was not trained to be factual or true representations of people or events." Stability lists the intended uses as artworks and design, educational or creative tools, and research on generative models — nothing else.

Technical Specifications

• Announced: October 22, 2024 (Large, Large Turbo); October 29, 2024 (Medium)
• Architecture: Multimodal Diffusion Transformer (MMDiT); MMDiT-X for the Medium variant
• Notable design choice: Query-Key Normalization integrated into the transformer blocks
• Parameters: 8.1B (Large); 2.5B (Medium); Large Turbo is a distillation of Large
• Output resolution: Up to 1 MP (Large); 0.25–2 MP (Medium). Stability publishes no resolution figure for Large Turbo.
• Sampling: Standard multi-step for Large and Medium; 4-step distilled for Large Turbo
• VRAM: 9.9 GB for Medium, excluding text encoders. Stability has not published figures for Large or Large Turbo.
• License: Stability AI Community License
• Weights: Hugging Face — inference code at Stability-AI/sd3.5

Use Cases

• Creative and concept art: Illustration, concept design, and reference material where prompt adherence matters more than raw throughput (SD 3.5 Large).
• High-volume generation: Batch and interactive workloads where four-step sampling makes the economics work (SD 3.5 Large Turbo).
• Local and offline generation: Running entirely on a consumer GPU with no API dependency (SD 3.5 Medium).
• Fine-tuned and domain-specific models: Adapting the base weights to a house style, product catalog, or character set.
• Product and marketing imagery: Generating on-brand visual assets, with outputs owned by the user under the Community License.
• Research and derivatives: Open weights make SD 3.5 a standard baseline for academic and independent diffusion research.

Performance / Benchmarks

Stability AI publishes no benchmark numbers for Stable Diffusion 3.5. All three model cards carry a "Model Performance" heading, and under all three the entire body of that section is a single sentence: "See blog for our study about comparative performance in prompt adherence and aesthetic quality." The blog it points to contains no win rates, no Elo, no percentages — only vendor-reported qualitative claims: Large "leads the market in prompt adherence and rivals much larger models in image quality"; Large Turbo "offers some of the fastest inference times for its size, while remaining highly competitive in both image quality and prompt adherence"; Medium "outperforms other medium-sized models, offering a balance of prompt adherence and image quality."

There is therefore no vendor benchmark table to reproduce here. What exists is independent arena data, and after twenty months it is unkind. Note that arena.ai (the former LMArena) and Artificial Analysis are different organisations running different Elo scales; their numbers are not comparable to each other.

arena.ai text-to-image leaderboard (fetched 2026-07-08; vote cutoff July 5, 2026)

SD 3.5 Large ranks #71 of 72 ranked models, on 23,396 of the board's 5,656,156 votes. Exactly one model sits below it. arena.ai carries no row for SD 3.5 Medium or SD 3.5 Large Turbo.

Artificial Analysis text-to-image leaderboard (fetched 2026-07-08)

Of 149 ranked models. For scale on this same board, Z-Image Turbo is #69 at 1,104, SDXL 1.0 is #145 at 875, and Stable Diffusion 1.5 is last at #149 with 660.

Both boards say the same thing in different units: SD 3.5 is not competitive on output quality against models released since, including open-weights ones. Its own 49% win rate on Artificial Analysis is a coin flip against the field.

The measure Stability still wins

Elo is not the only thing worth counting, and on the other one no lab is close. These are trailing-30-day download counts from the Hugging Face API, ranked within the text-to-image pipeline tag, queried September 6, 2026:

Read that table carefully, because it does not say what a Stability press release would want it to say. The two checkpoints ahead of everything are SDXL 1.0 (July 2023) and Stable Diffusion 1.5 (October 2022) — the U-Net generation, not the MMDiT one — and third place is an SD 1.5 fine-tune. SD 1.5's canonical repository is not even under the stabilityai organisation any more; it lives at the community-maintained stable-diffusion-v1-5 org. SD 3.5 Large, the flagship this page is about, draws under 5% of SDXL's downloads and sits outside the Hub's top fifty.

Treat the SD 3.5 rows as noisy rather than as a trend line: between the July 8 and September 6 queries, Medium fell from 617,267 to 152,888 while Large rose from 29,486 to 77,921. A trailing-30-day window on a mid-volume repository moves on mirror traffic and CI, not on adoption. The stable signal is the gap between the U-Net generation and the MMDiT one, and that gap is unchanged.

What that measures is fine-tuning demand and ecosystem lock-in — the LoRAs, ControlNets, and ComfyUI graphs built on the older architectures over three years — not image quality. The Stable Diffusion lineage still owns open-weights image generation in practice. Stable Diffusion 3.5 does not.

Limitations

• Revenue-capped free license: The Community License terminates once you or your affiliates exceed $1,000,000 in annual revenue, regardless of whether that revenue comes from the model. Above the threshold you must request an enterprise license, which Stability may grant at its discretion. Commercial users are also required to register with Stability AI.
• Resolution ceiling on Large: The flagship generates up to 1 megapixel — below Medium's 2 MP ceiling.
• Turbo trades quality for speed: Four-step distillation cuts inference time but does not match the full Large model's fidelity.
• Text encoder overhead: The 9.9 GB VRAM figure for Medium excludes text encoders, so real-world memory requirements are higher.
• Ageing against closed competitors: SD 3.5 dates from October 2024, and Stability has shipped no image-model successor since. Closed image models and the image capabilities folded into frontier multimodal systems have continued to advance.
• Content safeguards: The models carry safety filters that can be restrictive on legitimate borderline prompts.

Pricing & Access

Self-hosting

Free under the Stability AI Community License, subject to the $1M annual-revenue ceiling. Weights are on Hugging Face and inference code is at Stability-AI/sd3.5.

Direct model links:

• stable-diffusion-3.5-large
• stable-diffusion-3.5-medium

Hosted access

Stability lists these routes for using SD 3.5 without self-hosting:

• Stability AI API — see the platform pricing page for current credit costs
• Replicate
• Fireworks AI
• DeepInfra
• ComfyUI

Commercial licensing

• Under $1M annual revenue: Free for commercial use under the Community License, with registration.
• Over $1M annual revenue: The Community License terminates; an enterprise license must be requested from Stability AI.
• Output ownership: Users retain ownership of the images they generate.

Ecosystem & Tools

• Stability AI Platform: Official API access, credits, and commercial license management.
• Hugging Face: Weights for all three variants, plus community fine-tunes and LoRAs.
• Stability-AI/sd3.5 on GitHub: Reference inference code.
• ComfyUI: The dominant node-based workflow tool for local SD 3.5 generation; also listed by Stability as a hosted access route.
• Replicate, Fireworks AI, DeepInfra: Third-party hosted inference partners named in Stability's launch announcement.

What Stability shipped in 2026 instead

The gap since October 2024 is specific to the image line, and reading it as company-wide silence gets the picture wrong. Verified against Stability's own news feed on September 6, 2026:

• Stable Audio 3.0 (May 20, 2026): A four-model audio family — Small SFX, Small, Medium, Large — trained, Stability says, on fully licensed data. Small SFX, Small and Medium ship as open weights on Hugging Face under the same Community License as SD 3.5; Large is API and enterprise-self-host only. Medium and Large generate tracks past six minutes.
• Brand Studio (April 8, 2026): A managed creative-production platform for brand teams. Its Curated Model Routing picks per task across Stability's own models and third-party ones — Stability's commercial front end does not treat SD 3.5 as the only answer.
• Stable Audio DAW plugin (August 18, 2026): Stable Audio wired into a digital audio workstation rather than a web form.
• $76M Series B (August 25, 2026): New investors include Electronic Arts, Sony Music Group, Universal Music Group, Warner Music Group and AMD Ventures, taking total funding to $232M.

The pattern is a company that has moved its new-model effort to audio and its revenue effort to enterprise tooling. For an SD 3.5 adopter the practical read is that the weights are not going anywhere, and neither is a successor.

Community & Resources

• Introducing Stable Diffusion 3.5 - The official announcement, with variant specs and licensing terms
• Stable Image - Stability's current image model lineup
• Stability AI Community License - Full license text, including the $1M revenue clause
• Stability AI News & Updates - Where any successor model would be announced
• Stability AI Platform Pricing
• Stability AI on X