Module 22.1 closed on frontier language vendors. Volume 22 now switches modality: Image AI as an ecosystem catalog—who you buy, license, or self-host when stills (and short open video) are the job. Stability AI is the open-weight pole of that aisle: Stable Diffusion, SDXL, Stable Video Diffusion, plus the ControlNet / LoRA / UI stack that made local generation a product.
Volume 16.3/16.4 already scored these SKUs as models. Volume 17 is the textbook: diffusion, DDPM, latent space, SD wiring, SDXL tensors, ControlNet, ComfyUI. This 22.2 lecture is the vendor view: licenses, Stability Platform vs self-host, when Stability loses to FLUX or closed SaaS. Do not invent unverified UNet internals here—open Vol. 17 for that.
Learning Objectives
By the end of this lesson, students should be able to:
- Place Stability AI as the open-weight still/video lab behind SD, SDXL, SVD, and the ControlNet-era ecosystem (not a single checkpoint).
- Separate self-host weights + RAIL-class licenses from Stability Platform / DreamStudio-class paid APIs.
- Score Pricing / Strengths / Weaknesses qualitatively against closed Image AI (Midjourney, Firefly) without fake dollar figures.
- Run a minimal
diffusersstill (and know SVD is I2V, not Sora). - Point to Vol. 16.3/16.4 catalogs and Vol. 17 diffusion/ControlNet/ComfyUI for mechanism.
- Know when to leave Stability for FLUX (prompt/text), Firefly (indemnity), or Midjourney (look).
Stability AI is the company that popularized open latent diffusion stills: the Stable Diffusion line (1.x/2.x), SDXL (native ~1024, dual text encoders, optional refiner), later SD3-class SKUs, and Stable Video Diffusion (SVD)—open image-to-video with temporal layers on an SD-class backbone. ControlNet is a community/research adapter (Zhang et al.) that became inseparable from the SD ecosystem: pose, Canny, depth as extra conditioning. Product surfaces include downloadable weights (Hugging Face), Stability Platform APIs, and third-party UIs (ComfyUI, Automatic1111). ControlNet is not a Stability-owned model card in the same sense as SDXL—treat it as ecosystem, not a Stability SKU you “buy from DreamStudio.”
Pricing, Strengths & Weaknesses
Vol. 22 requires a vendor scorecard on every company. Pricing is a shape (GPU hours vs credits vs seats)—never freeze a blog’s dollar rate. Licenses travel with the checkpoint file, not the brand name.
| Axis | Stability AI (weights + optional API) | Closed still SaaS (MJ / Firefly / Ideogram) |
|---|---|---|
| Pricing | Self-host: your GPU / cloud instance hours. Stability Platform: usage or credit plans on current model ids. Community fine-tunes may be free to download and expensive to run. | Subscription + credit meters; no VRAM bill, but every still burns plan quota. |
| Strengths | Open weights; LoRA / DreamBooth / ControlNet; on-prem and air-gap; huge checkpoint zoo; SVD for short open I2V; diffusers + ComfyUI as engineering lingua franca. | Zero GPU ops; stronger out-of-box look or type (vendor-dependent); enterprise ToS / indemnity on some plans. |
| Weaknesses | You own VRAM, samplers, safety, and license reading. Vanilla SD 1.5/XL text and prompt fidelity lag FLUX and Ideogram. RAIL / community licenses are not Apache-by-default. SVD is short I2V, not closed T2V cinema. | No self-host; weak or no ControlNet-class lock; ToS and queues; harder to fine-tune a SKU. |
The Stack You Actually Buy
Checkpoints
- SD 1.5 / 2.x — still workhorses
- SDXL base + optional refiner
- SD3 / later Stability still SKUs
- SVD / SVD-XT I2V weights
Control & adapters
- ControlNet pose / Canny / depth
- LoRA / DreamBooth
- IP-Adapter / inpaint / outpaint
- Vol. 17 teaches the tensors
How you run it
diffusersin your service- ComfyUI graphs / A1111 lab
- Stability Platform HTTP jobs
- Replicate / HF Inference as a host
Minimal diffusers Still (SDXL)
Pin the Hugging Face id. SD 1.5 uses StableDiffusionPipeline; XL uses the XL pipeline. Samplers and CFG are Vol. 17—here you only need: load, prompt, steps, save. Confirm license on the card before commercial serving.
When Stability Is the Right Vendor
Pick Stability / SD-family when
- On-prem, VPC, or cost-per-image at scale on your GPUs
- You will train LoRAs on a product SKU
- ControlNet must lock pose, edges, or depth
- Short open I2V (SVD) is enough; no SaaS ToS
- ComfyUI graphs are the production contract
Move on when
- Prompt + short text quality → FLUX
- Opinionated cinematic look, no GPU → Midjourney
- In-image lettering → Ideogram
- Creative Cloud + indemnity story → Firefly
- Marketer templates → Canva
Safety checkers shipped with some diffusers pipelines are not a legal department. CreativeML Open RAIL-M / OpenRAIL++-M restrict certain uses. A Civitai fine-tune can mix licenses. SVD cards are often research/community-style—not Apache-2 by default. Read the file you load.
“Stability AI is one model called Stable Diffusion, ControlNet is a Stability product, and SVD is open Sora.” Stability is a lab + license family. SD, SDXL, SD3, and SVD are different checkpoints. ControlNet is an ecosystem adapter you pair with SD-class denoisers (Vol. 17). SVD is short open image-to-video (14/25 published frames in the original releases)—see 16.4 SVD—not closed text-to-video cinema. Vol. 16 catalogs the SKUs; Vol. 17 opens the denoiser; this lecture only tells you when the vendor is Stability.
Knowledge Check
- Short Answer: What is Volume 22.2’s job relative to Vol. 16.3 and Vol. 17 for Stability? Answer: 22.2 is the vendor/ecosystem buy; 16.3/16.4 catalog SKUs; Vol. 17 teaches diffusion/ControlNet/ComfyUI mechanism.
- True/False: ControlNet is the same checkpoint as SDXL and is sold only via Stability Platform. Answer: False—ControlNet is an ecosystem adapter; SDXL is a Stability-line base; you typically self-host both.
- Multiple Choice: SVD is best described as: (a) open minute-long T2V like Sora, (b) open I2V with temporal layers on an SD-class backbone, (c) a Midjourney video mode. Answer: (b).
- Short Answer: Name two Stability pricing shapes (no dollar amounts). Answer: Self-host GPU/cloud hours, and Stability Platform / API credits or usage plans.
- True/False: RAIL / community licenses mean unrestricted commercial use of any Civitai fine-tune. Answer: False—read each card; RAIL has use restrictions; fine-tunes may differ.
- Multiple Choice: Best reason to stay on Stability vs Firefly: (a) Discord grids, (b) self-host + ControlNet/LoRA + on-prem, (c) C2PA is automatic on every Civitai PNG. Answer: (b).
- Short Answer: Which Vol. 17 lectures should a student open after this vendor page? Answer: Diffusion/DDPM, latent space, SD, SDXL, ControlNet, ComfyUI (and SVD context from 16.4).
- True/False: Vanilla SD 1.5/XL typically beats Ideogram on in-image lettering. Answer: False—Ideogram is the typography specialist.
- Multiple Choice:
StableDiffusionXLPipelineis: (a) a Midjourney client, (b) a Hugging Face diffusers wrapper for SDXL stills, (c) the SVD video codec. Answer: (b). - Short Answer: When do you graduate from Stability stills to FLUX in this module? Answer: When prompt fidelity and short readable text matter more than the SD/ControlNet/LoRA zoo—next lecture.
Key Takeaways
- Stability AI is the open-weight still/video vendor: SD, SDXL, SVD, plus a ControlNet/LoRA/ComfyUI ecosystem.
- Pricing is GPU hours and/or API credits—not a fake list price; licenses live on the checkpoint.
- Vol. 16.3/16.4 catalog the SKUs; Vol. 17 teaches how they denoise.
- SVD is short open I2V, not closed T2V cinema.
- Next vendor: Black Forest Labs (FLUX).
Lab: Same product brief on SDXL (diffusers or ComfyUI) vs a closed still (Firefly or Midjourney if licensed). Score prompt lock, text, and whether ControlNet would have helped. No fake FID.
Whiteboard: Vendor vs catalog vs mechanism: 22.2 Stability → 16.3 SD/SDXL + 16.4 SVD → 17.1 diffusion/ControlNet. Draw checkpoint / adapter / UI as three boxes.
Recap: Stability AI is the open latent-diffusion vendor—self-host or API, RAIL not Apache-by-default, ControlNet in the ecosystem, SVD for short I2V. Continue with Black Forest Labs (FLUX).