Recap and final exam

Ten modules to go from a bare diffusers call on a pure-noise latent to a print-ready, provenance-manifested BoisClair catalog. Here is the course condensed, then the threads that ran through it, then the exam.

The course at a glance

Module	The essential point
1. Latent diffusion components	Autoencoder, U-Net, scheduler; text encoder conditions cross-attention; SDXL trained natively at 1024
2. Positive and negative prompts	Five slots (subject, medium, composition, lighting, style); token budget matters; negative prompts push away from named failure modes, they do not compose logic
3. Steps, guidance, seed	25–40 steps, guidance 6–9 for SDXL, DPM++ 2M Karras as default; seed is a first-class parameter — log it, sweep it, freeze it
4. Text-to-image and image-to-image	Denoising strength 0.3–0.5 for on-brand variants, 0.6–0.75 to reinvent background; do not chain img2img on img2img
5. Inpainting and outpainting	Dedicated fine-tuned checkpoint, feathered mask, high strength inside; outpaint in stages with mask overlap
6. ControlNet	Structural signal (Canny, depth, OpenPose); weight 0.4–0.7; end control before final steps to avoid over-constrained look
7. LoRA and textual inversion	20–40 captioned images, rank 4–8, ~2500 steps, trigger word, adapter weight 0.6–0.9; validate outside the style too
8. Upscaling and face restoration	Native 1024 → img2img at 2048 strength 0.25 → Real-ESRGAN x2; GFPGAN or CodeFormer for faces before the final upscale
9. Compute and memory	fp16 always, attention slicing and VAE tiling below 12 GB, CPU offload below 8 GB; resolution scales cost quadratically
10. Rights, provenance, use	OpenRAIL is a license, LoRAs have their own; no third-party logos or identifiable faces without consent; per-image provenance manifest

The threads running through the course

Everything is a latent trajectory. Text-to-image, image-to-image, inpainting, outpainting, ControlNet, LoRA-modulated generation, upscaling — all are variations on "start from a latent, denoise for N steps, decode with the VAE". Understanding that single frame makes every new pipeline in the ecosystem legible without having to relearn a new mental model.

Choose the smallest tool that solves your problem. A color variant is image-to-image at strength 0.4, not a new LoRA. A local retouch is inpainting, not a new full generation. A pose lock is ControlNet, not a longer prompt. Reaching for the biggest tool first is the fastest way to spend a week debugging a five-minute problem.

Reproducibility is a discipline, not a feature. Seed, model version, LoRA versions, scheduler, steps, guidance, denoising strength, ControlNet weights, prompt, negative prompt — logged with every image you keep. Without that log, "generate 20 more just like this one" becomes an unsolvable request three months later, and the C2PA provenance manifest of module 10 is impossible to build honestly.

Quality lives in the small numbers. Feathering a mask by 6 pixels, dropping guidance from 8 to 6.5, lowering ControlNet weight from 0.7 to 0.5, capping LoRA weight at 0.8, running a low-strength refine pass at 2048 — none of these look big on the parameter list. Together they are the difference between a demo and a catalog that ships.

The decision tree

A working shorthand of "which tool for which edit":

• New subject, no reference image → text-to-image, sweep the seed.
• Same subject, new material or color → image-to-image, strength 0.3–0.5.
• Same subject, new ambience → image-to-image, strength 0.6–0.75.
• Local detail wrong → inpainting, mask feathered, strength 0.9.
• Frame too tight for the format → outpainting, in stages with overlap.
• Composition or pose must be exact → ControlNet (depth, edges or pose), weight 0.5, end at 0.8.
• Every future image must look "like BoisClair" → LoRA on 20–40 curated images.
• Small motif or logo shape must recur → textual inversion, one token.
• Print or poster resolution needed → refine at 2× with strength 0.25 then Real-ESRGAN x2.
• Face looks plasticised → GFPGAN or CodeFormer before the final upscale.
• Machine OOMs → half precision, attention slicing, VAE tiling, then CPU offload.

The final exam

The exam has 40 questions covering all ten modules: the three components of latent diffusion, prompt structure and negative-prompt reality, step and guidance ranges, scheduler families and their reproducibility properties, seed as a first-class parameter, denoising strength as the master image-to-image dial, dedicated inpaint checkpoints and mask feathering, outpainting in stages, ControlNet weights and start/end scheduling, multi-ControlNet combination, LoRA dataset size and rank, trigger words and adapter weight, upscaling ladders and face restoration order, fp16 and attention slicing on tight VRAM, licenses and personality rights, provenance manifests.

Several questions present situations to diagnose: an image that looks plastic and over-baked at guidance 12, an image-to-image chain that has drifted the color palette by the fourth pass, an outpainting that produced a visible vertical seam at the mask boundary, a ControlNet weight that suppressed the material of the chair, a LoRA that leaked catalog watermarks into unrelated outputs, an OOM on the VAE decode after a two-minute generation. Judgement is what gets assessed, not memorized parameter names.

On success, your certificate of completion is issued immediately; its number can be verified by any third party on the platform.

Before you start

Re-read the decision tree above and, for each entry, ask yourself "what symptom would tell me I picked the wrong tool?". If you can explain why chained image-to-image drifts color, why a rank-16 LoRA on 20 images overfits, why guidance 12 looks plastic, and why a provenance manifest matters even for a small brand, you are ready. Good luck!

Final exam

Ready to validate this course?

40 questions drawn at random from the course bank · passing score 70% · verifiable PDF certificate issued immediately on success.

Start the exam

You need to be signed in to your InSkillML account with an active subscription. You can also start the exam from My courses.