[enhancement]: Need a download retry feature in Model Manager

[NiceDragon8066]

Is there an existing issue for this?

• I have searched the existing issues

Contact Details

pnjdutta464@gmail.com

What should this feature add?

I would like the InvokeAI team to include a feature in the Model Manager downloader code to keep trying to redownload at least five times in succession if the download fails or hangs for any reason. Downloading large models can take a very long time. Downloads can fail at any point due to ISP issues, power failure, host server problems or client-side problems. I have had InvokeAI model manager simply sit there doing absolutely nothing after a large model download failed anywhere between 0% to 100%. Model downloads failed after 9 or 10GB download of a 12GB model, and there is simply no way to get that download restarted! Restarting InvokeAI simply discarded the downloaded data with a message about clearing hanging folder temp-data etc. That's unacceptable. The Wan2GP app restarts its download upon restarting the app; why can't InvokeAI do the same?

Alternatives

The only alternative that reliably works so far is to manually download the model file and ditch the InvokeAI model manager completely for downloads.

Additional Content

No response

[DustyShoe]

Out of curiosity, does it work in any other tool?

[NiceDragon8066]

Replying to @DustyShoe - Yes, there are other platforms which retry failed or suspended downloads on their own. I already mentioned Wan2GP in my enhancement request above. If a download fails or hangs, closing and restarting the Wan2GP app resumes the download. Unlike InvokeAI Wan2GP does not discard already downloaded partial file. Even downloads in Chrome or Edge can be restarted or resumed. Where I am living, there is no such thing as unlimited highspeed internet data. I have to buy data in 50GB blocks, and I typically use around 1000GB per month. The data InvokeAI discards is essentially money from my pocket wasted.

[DustyShoe]

I'm sorry, meant to ask in another issue report

[DustyShoe]

The point is valid.
I left downloading of Klein 4b overnight just to find out that download is stalled. Cancelling and restarting creates new temp folder with unique name leaving old one dangling in models folder.
I see few points of improvements:

1. Being able to download failed parts if model consists of multiple files without creating new folder.
2. Delete folder with previous download if session is unrecoverable.

(3.) Explore possibility to implement manager that makes possible to continue stalled halfway download at least on large files.

[NiceDragon8066]

Well, do whatever you feel is suitable for your user base. Based on your unsatisfactory response to my bug report issue #8839, stating that my GTX 1650 wasn't supported, I got rid of my InvokeAI installation completely. So, it doesn't matter to me anymore. That response came as a big surprise as the app installer/launcher gives options for newer and older GPUs including GPUs older than 30xx or 20xx - that covers 10xx. Also, InvokeAI support archive shows how to run the app on low VRAM computers through setup optimization. Further I have successfully used the app to run large models and encoders for SDXL, FLUX and Z-IMAGE (by using low resolution generations), so I don't buy your excuse that my 4GB VRAM and 32GB RAM computer cannot run the app with large models due to memory "fracturing." Best you could've said that you didn't know how to solve the problem. Anything can be done with some creativity and persistence. Demanding more and more memory isn't a practical solution. Such apps are destined to garbage pile. I will stick with WebUI, ComfyUI and Wan2GP. Have good day.

…

On Thu, Feb 5, 2026, 12:16 PM DustyShoe ***@***.***> wrote: *DustyShoe* left a comment (invoke-ai/InvokeAI#8840) <#8840 (comment)> The point is valid. I left downloading of Klein 4b overnight just to find out that download is stalled. Cancelling and restarting creates new temp folder with unique name leaving old one dangling in models folder. I see few points of improvements: 1. Being able to download failed parts if model consists of multiple files without creating new folder. 2. Delete folder with previous download if session is unrecoverable. (3.) Explore possibility to implement manager that makes possible to continue stalled halfway download at least on large files. — Reply to this email directly, view it on GitHub <#8840 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BYUOHLL65GVWAWOMETOMVJT4KLRNLAVCNFSM6AAAAACT55NK36VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTQNJRGM4DCOBWGU> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[DustyShoe]

Sinse the issue is not relevant anymore, feel free to close it.
Otherwise i'll do it myself, becase conversation went off topic here.
Best regards.