A checkpoint is the “main brain” file that generates the image. Everything else (LoRA, ControlNet, etc.) sits on top of a checkpoint.

This guide focuses only on checkpoints:

  • Base models (the foundations)
  • Trained models (fine-tunes) vs merge models
  • The most commonly used base families + their typical use-cases
  • Practical pros/cons + notable restrictions/flaws

1) Base checkpoint vs fine-tuned checkpoint

Base checkpoint

A general-purpose foundation released by a lab/vendor (or a large community release). Examples: SD 1.5, SDXL 1.0, Stable Diffusion 3.5, FLUX.1 [dev].

Why you care: the base decides capabilities and ecosystem support (tools, LoRAs, ControlNets, workflows).

Fine-tuned checkpoint (trained model)

A base checkpoint that has been further trained on a dataset to specialize style/content (e.g., “anime portraits”, “product renders”, “cinematic realism”). Example: Pony Diffusion v6 XL is an SDXL finetune.

Why you care: you get “the look” faster, often with shorter prompts.

2) Trained models vs merge models

Trained model (fine-tune)

How it’s made: additional training updates the weights using examples (gradient descent). Typical result: more coherent “personality” and consistent behavior.

Pros

  • More reliable style/subject bias
  • Often better consistency across prompts
  • Usually easier prompting (less wrestling)

Cons

  • Can overfit (repeating faces/compositions)
  • Strong biases (e.g., “everything becomes glossy portraits”)
  • Quality depends heavily on dataset curation

Merge model

How it’s made: combines two+ checkpoints (weight blending) without extra training. Typical result: a “cocktail” of traits (e.g., realism from one + lighting from another).

Pros

  • Fast to create, lots of experimentation
  • Can mix complementary strengths

Cons

  • Can be less stable (odd artifacts, inconsistent prompt response)
  • “Best of both” is not guaranteed; sometimes you get “worst of both”
  • Lineage/licensing can get messy if merged components have restrictions

3) The most used base families (and what they’re good at)

Below are widely used foundations or foundation-like families you’ll see constantly across tools and community checkpoints.

A) Stable Diffusion 1.5 (SD 1.5)

Best for: lightweight local generation, huge ecosystem, tons of fine-tunes/LoRAs.

Pros

  • Runs on modest GPUs
  • Massive community library (styles, characters, niches)

Cons / common flaws

  • Weaker at text-in-image and complex compositions than newer bases
  • Anatomy/hands often need extra help (inpainting, ADetailer-like workflows)

License note: released under CreativeML Open RAIL-M.

B) Stable Diffusion 2.1 (SD 2.1)

Best for: “cleaner” general image generation compared to 1.5, some workflows prefer it.

Pros

  • Solid general base; can feel less “muddy” than many old 1.5 blends

Cons

  • Smaller community ecosystem than 1.5/SDXL
  • Many users find it pickier with prompts than 1.5-style finetunes

License note: commonly distributed under CreativeML Open RAIL++-M.

C) Stable Diffusion XL 1.0 (SDXL)

Best for: higher resolution, better prompt understanding, better typography than SD1.5 in many cases.

Pros

  • Strong general-purpose base with modern “look”
  • Great for: portraits, environments, design-ish images, better composition

Cons

  • Heavier GPU/VRAM demands than SD1.5
  • Many SDXL finetunes expect SDXL-specific workflows (refiners, SDXL VAEs, etc.)

License note: CreativeML Open RAIL++-M (official release).

D) SDXL Turbo / SD-Turbo (distilled “fast” variants)

Best for: speed, quick ideation, real-time-ish exploration.

Pros

  • Very fast generation (great for sketching ideas)

Cons

  • Often lower fidelity than full SDXL at the same resolution
  • Less “room” to refine details; can look a bit “compressed”

License note: SD-Turbo points to Stability’s licensing terms; commercial terms depend on the license path.

E) Stable Diffusion 3.5 (SD 3.5 suite)

Best for: modern prompt-following and high quality on consumer hardware (where supported).

Pros

  • Designed as a newer generation; positioned as customizable and consumer-hardware friendly

Cons

  • Tooling/workflow compatibility can be more variable than SDXL/1.5 depending on your stack

License note: Stability states SD 3.5 models are free for commercial + non-commercial use under the Stability AI Community License.

F) FLUX.1 [dev] (Black Forest Labs)

Best for: high-quality generations and strong modern results (where supported).

Pros

  • Strong general image quality and “modern” rendering (popular momentum in 2024–2025+)

Cons / restrictions

  • The model is under a “Non-Commercial License” for the weights, but the license explicitly states users can use outputs broadly (with caveats) and forbids using outputs to train/fine-tune a competing model.
  • Heavier compute requirements than SD1.5; best experience depends on your tooling

G) “Pony” (as a foundation-like community base: Pony Diffusion XL)

Best for: anime/cartoon/anthro ecosystems, Danbooru-tag-style prompting culture, huge LoRA scene around it.

Pros

  • Very strong community “content + LoRA” ecosystem in its niche
  • Often excellent for stylized characters and illustration workflows

Cons / quirks

  • Prompting style can differ from “plain SDXL”: many users lean into tag-like prompting
  • Can bias outputs toward the dataset’s dominant styles

Evidence note: Pony Diffusion v6 XL is explicitly an SDXL finetune in its model card.

H) “Anything” (SD1.5 anime finetune family, e.g., Anything V5)

Best for: classic SD1.5 anime look, fast stylized portraits/characters.

Pros

  • Easy to get anime-style results with short prompts
  • Lightweight (SD1.5 base)

Cons

  • Less flexible outside its core aesthetic vs newer SDXL-based anime finetunes

Evidence note: Anything V5 is described as an anime-focused fine-tuned SD checkpoint.

I) “Illustrious XL” (SDXL illustration-focused checkpoint family, OnomaAIResearch)

Best for: high-resolution illustration/anime-style images (notably native 1536×1536) with a mix of natural-language + Danbooru/tag prompting.

Pros

  • Native 1536px support (and wide resolution range without extra hi-res tricks)
  • Works well with plain English prompts, tags, or both (hybrid prompting)
  • Strong as a foundation for training/using add-ons (the model card highlights compatibility with common extensions like LoRA/ControlNet)

Cons

  • Version behavior can vary; the authors publish a v2.0 “STABLE” checkpoint specifically for more stable generation behavior
  • Knowledge cutoff noted on v1.0 (trained July 2024; knowledge up to June 2024)
  • Licensing differs by release: v1.0 is marked sdxl-license, while v2.0 is marked creativeml-openrail-m—so always check the exact version you download before commercial use/distribution

Evidence note: On Hugging Face, Illustrious XL v1.0 is described as an SDXL-based model with native 1536×1536 plus NLP + tag-based prompting, and v2.0 is presented as a more stabilized checkpoint.

4) Practical selection guide (fast)

  • I want maximum compatibility + huge community assets: SD 1.5 or SDXL.
  • I want a modern all-rounder at higher quality: SDXL 1.0 (then pick a finetune like a realism/art variant).
  • I want speed for brainstorming: SDXL Turbo / SD-Turbo.
  • I want high-res illustration/anime with hybrid prompting (English + tags): Illustrious XL.
  • I want anime/stylized character pipelines (tag culture + big LoRA ecosystem): Pony (SDXL finetune).
  • I want classic lightweight SD1.5 anime aesthetics: Anything-style (SD1.5 finetune).
  • I want strong modern results and my tooling supports it (but check license constraints): FLUX.1 [dev].