.ai

the best open source ai models

Guides

Rankings

FAQ

Upload

Which AI image models to try first, and why: a ranked starting list

Q: Too many choices on the family page

Fine-tunes of a popular base number in the hundreds. Sort by hot, pick the top checkpoint, ignore the rest for a week.

Q: Video model produces a slideshow or a still

Motion prompt missing or the frame count too low. Describe the motion, not the scene; 81 frames on Wan 2.2; keep the camera instruction to one movement.

Topic: Models

By Captain

Published 2025-09-10

Seven models for a first week, chosen for forgiveness rather than hype: one SDXL photo fine-tune, one anime fine-tune, Flux dev, a fast turbo model, an inpainting model, a video model and an upscaler, each with the reason, the memory it needs and the prompt style it expects.

Overview

The hottest model of the week is not always the best first model. New releases need the newest tools, the biggest GPUs and prompts nobody has figured out yet. A good first model is forgiving: it looks good with a short prompt, it runs on an 8 GB card or on the cloud partners, and it has hundreds of

LoRAs

and example images to learn from. This list is built on that rule, using the download and run figures this site collects.

The list is one model per job. Image generation is not one task but five or six (photo, illustration, editing, fast drafts, video, enlargement), and the model that is best at one is usually mediocre at another. Learn one per job, then explore the ranking pages for alternatives. The weekly ranking on this site shows what is replacing what; the family pages show the whole lineage.

Everything here is linked to its catalog page, where you will find the exact prompts of the example images, the file hashes, the memory requirements and a Run button for the two cloud partners. Try each on the cloud before downloading; a

checkpoint

is 6 GB for

SDXL

and 12 to 24 GB for

Flux

, and your taste will change within the week.

Reference

Name	Type	What it is
1. SDXL photo fine-tune	DreamShaper XL / Juggernaut XL / RealVisXL	Photo and general purpose. 1024x1024, CFG 5-7, 25-30 steps, 8 GB. Most forgiving of mediocre prompts; thousands of LoRAs.
2. Illustrious or Pony fine-tune	WAI-Illustrious / Hassaku XL / CyberRealistic Pony	Anime and stylised characters. Tag prompts (1girl, masterpiece, artist tags ). 8 GB. Knows thousands of characters by name.
3. Flux dev (fp8)	flux1-dev-fp8 / Flux Krea dev	Best prompt following and text rendering of the open models of its year. Sentences, guidance 3.5, 20-28 steps. 12 GB fp8, 8 GB with GGUF Q4.
4. A turbo / lightning model	SDXL Lightning / Z-Image Turbo / FLUX.2 Klein	Drafting. 4 to 8 steps, CFG 1-2, seconds per image. Use it to find a composition, then redo on the full model.
5. An inpainting model	DreamShaper 8 inpainting / Flux Fill / Qwen Image Edit	Fixing hands, swapping objects, extending canvases. Same prompt style as its family.
6. A video model	Wan 2.2 (5B or 14B) / LTX-2	Image to video and text to video. 16 GB locally for the small versions; the cloud partners run the large ones.
7. An upscaler	4x-UltraSharp / Real-ESRGAN / SwinIR	Enlarging finals 2x to 4x. Tiny files, run in seconds, part of every UI.

Try all seven without downloading 60 GB

Every model on this list is a click away in BitVector Prism, a web app with the models preinstalled. Work through the seven-day plan on a laptop or a phone, then decide what deserves space on your own disk.

Try BitVector

Step by step

Day 1: pick the SDXL photo fine-tune. Copy three example prompts from its page, change the subject, generate at 1024x1024, 28 steps, DPM++ 2M Karras, CFG 6.
Day 2: the anime fine-tune. Read the
Illustrious
and Pony tag guide on this site first; prompts here are comma-separated tags with quality tags in front.
Day 3:
Flux dev
. Write a full sentence, no
negative prompt
, guidance 3.5, 24 steps. Notice how it handles text in quotes and counted objects.
Day 4: the turbo model. Generate 20 drafts in the time one Flux image took; keep the compositions you like and recreate the best on Flux or SDXL with the same seed shape.
Day 5:
inpainting
. Take a Day 1 image with a bad hand, mask it, prompt "hand resting on the table", denoise 0.6. Read the img2img and inpainting guide.
Day 6: video. Take your best image and animate it with
Wan 2.2
image-to-video, a short motion prompt ("she turns her head and smiles, slow camera push-in"), 81 frames at 16 fps.
Day 7: upscale the three finals you like most with 4x-UltraSharp, then downscale to 2x in an image editor for crisp results.
After the week, open the weekly ranking and the family pages and start replacing one model at a time.

Examples

The same subject, three first models

SDXL photo: "a street musician playing accordion in a rainy Lisbon alley, warm shop lights, 35mm, candid" | 1024x1024 | 28 steps | CFG 6

Illustrious: "1boy, accordion, street, rain, night, warm lighting, from side, masterpiece, best quality, very aesthetic" | 832x1216 | 28 steps | CFG 5

Flux dev: "A candid 35mm photo of a street musician playing an accordion in a narrow rainy alley in Lisbon at night; warm light spills from a bakery window on the left." | 1024x1024 | 24 steps | guidance 3.5

PirateDiffusion commands for the week

/render <juggernaut-xl> a street musician playing accordion in a rainy alley, warm lights, 35mm /size:1024x1024

/render <wai-illustrious> 1boy, accordion, street, rain, night, masterpiece, best quality /size:832x1216

/workflow /run:wan22-i2v slow camera push in, he looks up and smiles (reply to your image)

Tips

Use the family badge on each card: a LoRA only works with checkpoints of the same family, so learn families before you learn model names.
The Favs heart on the cards keeps your short list in your browser; no account is needed and nothing is sent to us.
If the first results are ugly, change the model before the prompt. A wrong model needs a perfect prompt; a right model needs a short one.
Pony and Illustrious are both SDXL but prompt differently and share few LoRAs. Pick one for the first week.
Flux Krea
dev, Flux SRPO and the many realism LoRAs exist because plain Flux dev has a recognisable glossy look; the family page lists the fixes.
Z-Image Turbo
and
FLUX.2
Klein arrived after this list was written and are now better first fast models than SDXL Lightning; the weekly ranking stays current even when articles do not.

Describe the motion, not the scene; 81 frames on Wan 2.2; keep the camera instruction to one movement.

The whole list, one subscription

PirateDiffusion has the SDXL fine-tunes, Flux, the turbo models, inpainting, Wan video and the upscalers installed on Telegram. Unlimited generation on a fixed price plan, so the week of experiments costs the same as one image.

Try PirateDiffusion

Questions

Should I start with the newest model?

No. Start with the model that has the most LoRAs and examples in its family; newness buys little for a beginner. Move up once the basics are boring.

SD 1.5 is small and fast, why is it not on the list?

Its hands, faces and prompt following are a generation behind. It is still worth knowing for inpainting and for 4 GB cards, and the

SD 1.5

family page lists the best fine-tunes.

One model for everything?

Flux dev or

Qwen Image

come closest for images, Wan 2.2 for video. Specialists still win in their lane.

Where do these recommendations come from?

Downloads, runs and search interest collected by this site, plus the LoRA count per family, which measures how much a community has invested in a base.