SD and SDXL are the classic UNet latent-diffusion catalog. FLUX (Black Forest Labs, 2024+) is the next open-weight generation many teams actually deploy: a rectifier/flow-matching transformer family with a three-SKU license split—[schnell] Apache-2, [dev] non-commercial unless you buy a license, [pro] API-only.
Volume 17 has a dedicated FLUX lecture plus DDPM vs flow context, ControlNet-style control, and ComfyUI graphs. Here you learn to buy and license FLUX correctly. Closed photoreal APIs continue with Google Imagen.
Learning Objectives
By the end of this lesson, students should be able to:
- Name the FLUX.1 SKUs: schnell, dev, pro—and what each license allows.
- Score FLUX on the 16.3 rubric vs SDXL and DALL·E.
- Run a schnell
diffuserscall and know why few steps / low guidance are typical. - Treat FLUX as flow-matching / transformer, not “just another SDXL UNet.”
- Plan VRAM, quantization, and ComfyUI serving without inventing FLOPs.
- Point to Vol. 17 for architecture, ControlNet-like control, and production graphs.
FLUX.1 is Black Forest Labs’ text-to-image family. Architecturally it is closer to a multimodal diffusion transformer with flow matching than to SD 1.5’s UNet-DDPM. Product-wise it ships as schnell (fast, openly licensed Apache-2.0), dev (open weights but a non-commercial / research-style license unless BFL grants a commercial license), and pro (closed, API). Later BFL image SKUs may appear—always re-read the card; the three-way license pattern is the durable lesson.
Rubric: The License Split Is the Product
| Axis | FLUX.1 [schnell] | FLUX.1 [dev] | FLUX.1 [pro] |
|---|---|---|---|
| Open vs closed | Open weights + Apache-2.0 | Weights downloadable; license not Apache | Closed API |
| Prompt fidelity | Strong for a few-step model; slightly behind dev/pro on hard briefs | Strong literal following among open-weight stills | Vendor’s best FLUX stills |
| Text rendering | Much better than SD 1.5/SDXL; still not a layout engine | Typically strongest of the open FLUX pair | Strong; Ideogram still specializes in posters |
| Licensing / commercial | Apache-2.0—commercially friendly for that checkpoint | Non-commercial unless you obtain a BFL commercial license | BFL/API ToS + paid usage |
| Latency | Few inference steps by design (often ~1–8) | More steps than schnell; heavier | Vendor queue; seconds-class stills |
Why Teams Switch from SDXL
Wins
- Prompt following without a prompt novel
- Readable short text more often
- Apache schnell for legal-simple serving
Costs
- VRAM: full FP16/BF16 is hungry
- dev license foot-guns in startups
- Smaller LoRA zoo than SD 1.5 (growing)
Vol. 17
- Flow matching vs DDPM
- ComfyUI FLUX graphs
- Control / LoRA ports
schnell via diffusers (Pattern)
HF id is commonly black-forest-labs/FLUX.1-schnell. Few steps and guidance_scale=0 (or very low) are typical for schnell because it was distilled for that regime—do not blindly copy SDXL’s 30 steps @ 7.5. Confirm the current card; quantization (FP8, GGUF, NF4) is how 24 GB cards survive, not a quality leaderboard.
Typical Use Cases
Pick FLUX when
- Open (or API) stills with better text/layout than XL
- schnell Apache serving is the legal win
- You will invest in ComfyUI + quant
- DALL·E fidelity without OpenAI lock-in
Move on when
- Poster typography is the product → Ideogram
- Indemnification → Firefly
- Tiny GPU / 1.5 LoRA museum → SD 1.5
- You accidentally used [dev] in a paid SaaS
“FLUX is open, so [dev] is fine in our commercial app.” Downloadable weights ≠ commercial license. Shipping [dev] in a paid product without a BFL commercial grant is a classic startup foot-gun. Use [schnell] (Apache), pay for [pro]/a commercial [dev] license, or pick another stack. Legal review of the exact Hugging Face license file is part of model selection—same discipline as RAIL on SD.
Knowledge Check
- Short Answer: Name the three FLUX.1 SKUs and one license fact each. Answer: schnell Apache-2; dev non-commercial unless licensed; pro closed API.
- True/False: FLUX [dev] weights being on Hugging Face means Apache-2 commercial use. Answer: False—dev is not Apache; commercial use needs a BFL license.
- Multiple Choice: schnell inference typically uses: (a) 50 DDPM steps @ CFG 12, (b) few steps and low/zero guidance, (c) Discord --stylize. Answer: (b).
- Short Answer: Why is FLUX often chosen over SDXL? Answer: Stronger prompt following and text/layout among open-ish still models.
- True/False: FLUX is just a larger SDXL UNet. Answer: False—transformer / flow-matching family, not classic SD UNet-DDPM.
- Multiple Choice: Best Vol. 17 follow-ups: (a) MCP transports, (b) FLUX + DDPM/flow + ComfyUI, (c) Whisper diarization. Answer: (b).
- Short Answer: What should a startup do instead of silently serving [dev]? Answer: Use schnell, buy a commercial license / pro API, or switch stacks.
- True/False: FLUX solves typesetting as well as Ideogram. Answer: False—better than SDXL, still not a poster specialist.
- Multiple Choice: VRAM for full BF16 FLUX is typically: (a) trivial on 8 GB, (b) a real constraint—quant/offload/API, (c) zero because it is CSS. Answer: (b).
- Short Answer: Which next catalog lecture is Google’s closed photoreal API? Answer: Google Imagen.
Key Takeaways
- FLUX is BFL’s flow-matching T2I family with a three-way license split.
- schnell = Apache + few steps; dev = weights ≠ commercial; pro = API.
- Expect better fidelity/text than SDXL, higher VRAM, growing ComfyUI support.
- Vol. 17 teaches the architecture; this lecture teaches the purchase order.
- Continue with Google Imagen.
Lab: If GPU allows, run schnell at 4 steps. Same prompt on SDXL. Compare readable signage and prompt extras (unwanted objects). Discuss [dev] license with a fake “we shipped it in prod” scenario.
Whiteboard: Three columns: schnell / dev / pro. Fill license, latency, and who owns GPUs. Arrow to Vol. 17 FLUX + ComfyUI.
Recap: FLUX is licensed in three SKUs—do not confuse downloadable with commercial. Continue with Google Imagen.