Your noob to expert guide in making great AI art.

[not_a_user_69]

I tried a workaround with prompt, added Solo.
But still please share other workaround. You can explain whenever you got time no rush, thanks.
Can you please explain how to use Embedding for Lazy Embedding in Invoke?

 

[NoTraceOfLuck]

NoTraceOfLuck I know I am troubling you, here is the Positive prompt
Detailed hair, detailed face, detailed body, detailed background, beautiful eyes, score_9, score_8_up, score_7_up, masterpiece, best quality, amazing quality, highly detailed, (best shadow), (absurdres), HDR, highres, absurdres, detailed face, detailed eye, face focus, (mature woman:1.2), 1girl, (full shot:1.5), hero view, (front view:1.4), long black hair, emerald eyes, seductive gaze, makeup, lipstick, sweat, medium breasts, milf, sexy, detailed skin, photorealistic lighting, female teacher attire, sexy teacher, office lady, open shirt, (shirt unbuttoned:1.2), cleavage, visible collarbone, tight skirt, blazer open, sitting on desk, (dominant pose:1.1), seductive expression, confident, professional attire, classroom, blackboard, indoor, school setting,

Please craete a tutorial by generating an image for 16:9 ratio and upscaling it.
If I create using default resolution or 9:16, upscaling it to 16:9 FHD doesn't give satisfactory output.
and when I create using 16:9 resolution character is created side by side losing alignment, is there anything which could be done?
Simpler steps cause to create multiple images for 16:9 i need a efficient method. I know I am troubling, but please help me here.

Or is there any Lora available in CivitAI to tackle it?

I can provide some tips but I don't really have any new information to add that isn't already elsewhere in this tutorial. There is one important thing I should stress: AI will often create bad outputs. Follow my tutorial on inpainting and controlnet to learn how to manually fix these images. There is nothing I can say that will allow you to create perfect images. For my game, I would say 90% of the images I produce require manual fixes, only about 10% come out correctly in one shot.

But here are some notes I will make based on what you've shared:

1. Your prompt is extremely long, and much of that is quality / detail tags. This likely negatively affects your outputs. Follow my tutorial on quality tags to understand the best quality tags for your model. If you're using Pony, you only need the score_9, score_8 etc... tags. Trim your prompt down as small as you can get it, then try again
   1. If your Negative prompt is also very long, trim that down too.
2. Invoke does not support the Civitai style of prompt weights, it uses a different syntax. Eg: Instead of (front view:1.4) you will use (front view)++++ in invoke. Switch all of these and you will get a better output
3. Weighting things very highly can negatively affect outputs. Try generating again without any weighted tags and then slowly add the weights back in.
4. The model you're using may not be trained properly for 16:9 aspect ratio. Try using different models.

For embeddings, you can download them with the model manager, then access them via the same menu where I showed how to access 'Trigger Words' (you can search the tutorial for that phrase to see a screenshot of the menu)

 

[nicod]

is it really that hard to render on Weak pc ?!
i tried to render it on 64x64 lowest resolution and i took for me hours to happen

MY Ram : 16 giga
GPU : Nvidia GTX 1650

or do i have anyissue with my setup

Log

Started Invoke process with PID 3756
[2025-10-13 15:24:19,368]::[InvokeAI]::INFO --> PyTorch CUDA memory allocator: cudaMallocAsync
[2025-10-13 15:24:46,665]::[InvokeAI]::INFO --> Using torch device: NVIDIA GeForce GTX 1650
[2025-10-13 15:24:47,736]::[InvokeAI]::INFO --> cuDNN version: 90701
[2025-10-13 15:24:52,462]::[InvokeAI]::INFO --> Patchmatch initialized
[2025-10-13 15:24:53,640]::[InvokeAI]::INFO --> InvokeAI version 6.8.1
[2025-10-13 15:24:53,640]::[InvokeAI]::INFO --> Root directory = E:\AI
[2025-10-13 15:24:53,644]::[InvokeAI]::INFO --> Initializing database at E:\AI\databases\invokeai.db
[2025-10-13 15:24:53,746]::[ModelManagerService]::INFO --> [MODEL CACHE] Using user-defined RAM cache s
ize: 8.0 GB.
[2025-10-13 15:24:54,605]::[InvokeAI]::INFO --> Invoke running on

You must be registered to see the links

(Press CTRL+C t
o quit)
[2025-10-13 15:25:21,261]::[InvokeAI]::INFO --> Emptying model cache.
[2025-10-13 15:31:53,048]::[InvokeAI]::INFO --> Executing queue item 6, session 405de59e-5f29-4025-bbb8
-a9d0c78e1dde
Fetching 17 files: 100%|██████████████████████████████████████████████████████| 17/17 [00:00<?, ?it/s]
Loading pipeline components...: 100%|███████████████████████████████████| 7/7 [06:26<00:00, 55.18s/it]
[2025-10-13 15:38:39,547]::[InvokeAI]::WARNING --> Loading 0.146484375 MB into VRAM, but only -798.25 M
B were requested. This is the minimum set of weights in VRAM required to run the model.
[2025-10-13 15:38:39,976]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model '58a740e8-df7b-41
c4-9729-3d4c07950e39:text_encoder' (CLIPTextModel) onto cuda device in 0.69s. Total model size: 469.44M
B, VRAM: 0.15MB (0.0%)
[2025-10-13 15:38:40,383]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model '58a740e8-df7b-41
c4-9729-3d4c07950e39:tokenizer' (CLIPTokenizer) onto cuda device in 0.01s. Total model size: 0.00MB, VR
AM: 0.00MB (0.0%)
[2025-10-13 15:38:53,995]::[InvokeAI]::WARNING --> Loading 0.634765625 MB into VRAM, but only -860.25 M
B were requested. This is the minimum set of weights in VRAM required to run the model.
[2025-10-13 15:38:54,021]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model '58a740e8-df7b-41
c4-9729-3d4c07950e39:text_encoder_2' (CLIPTextModelWithProjection) onto cuda device in 0.05s. Total mod
el size: 2649.92MB, VRAM: 0.63MB (0.0%)
[2025-10-13 15:38:54,030]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model '58a740e8-df7b-41
c4-9729-3d4c07950e39:tokenizer_2' (CLIPTokenizer) onto cuda device in 0.01s. Total model size: 0.00MB,
VRAM: 0.00MB (0.0%)
[2025-10-13 15:39:08,760]::[InvokeAI]::WARNING --> Loading 0.146484375 MB into VRAM, but only -860.25 M
B were requested. This is the minimum set of weights in VRAM required to run the model.
[2025-10-13 15:39:08,768]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model '58a740e8-df7b-41
c4-9729-3d4c07950e39:text_encoder' (CLIPTextModel) onto cuda device in 0.03s. Total model size: 469.44M
B, VRAM: 0.15MB (0.0%)
[2025-10-13 15:39:08,773]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model '58a740e8-df7b-41
c4-9729-3d4c07950e39:tokenizer' (CLIPTokenizer) onto cuda device in 0.00s. Total model size: 0.00MB, VR
AM: 0.00MB (0.0%)
[2025-10-13 15:39:09,141]::[InvokeAI]::WARNING --> Loading 0.634765625 MB into VRAM, but only -860.25 M
B were requested. This is the minimum set of weights in VRAM required to run the model.
[2025-10-13 15:39:09,159]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model '58a740e8-df7b-41
c4-9729-3d4c07950e39:text_encoder_2' (CLIPTextModelWithProjection) onto cuda device in 0.04s. Total mod
el size: 2649.92MB, VRAM: 0.63MB (0.0%)
[2025-10-13 15:39:09,167]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model '58a740e8-df7b-41
c4-9729-3d4c07950e39:tokenizer_2' (CLIPTokenizer) onto cuda device in 0.01s. Total model size: 0.00MB,
VRAM: 0.00MB (0.0%)
Fetching 17 files: 100%|███████████████████████████████████████████| 17/17 [00:00<00:00, 26516.61it/s]
Loading pipeline components...: 100%|███████████████████████████████████| 7/7 [07:00<00:00, 60.07s/it]
[2025-10-13 15:46:46,903]::[InvokeAI]::WARNING --> Loading 1.904296875 MB into VRAM, but only -860.25 M
B were requested. This is the minimum set of weights in VRAM required to run the model.
[2025-10-13 15:46:47,401]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model '58a740e8-df7b-41
c4-9729-3d4c07950e39:unet' (UNet2DConditionModel) onto cuda device in 0.69s. Total model size: 9794.10M
B, VRAM: 1.90MB (0.0%)
Fetching 17 files: 100%|████████████████████████████████████████████| 17/17 [00:00<00:00, 6716.58it/s]
Loading pipeline components...: 100%|██████████████████████████████████| 7/7 [31:14<00:00, 267.79s/it]
[2025-10-13 16:18:49,050]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model '58a740e8-df7b-41
c4-9729-3d4c07950e39:scheduler' (EulerDiscreteScheduler) onto cuda device in 0.16s. Total model size: 0
.00MB, VRAM: 0.00MB (0.0%)
100%|█████████████████████████████████████████████████████████████████| 30/30 [08:39<00:00, 17.32s/it]
Fetching 17 files: 100%|██████████████████████████████████████████████████████| 17/17 [00:00<?, ?it/s]
Loading pipeline components...: 100%|███████████████████████████████████| 7/7 [08:38<00:00, 74.09s/it]
estimate_vae_working_memory_sd15_sdxl: 9489612800
[2025-10-13 16:37:07,551]::[InvokeAI]::WARNING --> Loading 0.0 MB into VRAM, but only -5814.25 MB were
requested. This is the minimum set of weights in VRAM required to run the model.
[2025-10-13 16:37:07,558]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model '58a740e8-df7b-41
c4-9729-3d4c07950e39:vae' (AutoencoderKL) onto cuda device in 1.53s. Total model size: 319.11MB, VRAM:
0.00MB (0.0%)
[2025-10-13 16:37:23,653]::[InvokeAI]::INFO --> Graph stats: 405de59e-5f29-4025-bbb8-a9d0c78e1dde
Node Calls Seconds VRAM Used
sdxl_model_loader 1 0.048s 0.000G
sdxl_compel_prompt 2 438.360s 0.245G
collect 2 0.094s 0.009G
string 1 0.001s 0.009G
integer 1 0.002s 0.009G
core_metadata 1 0.166s 0.009G
noise 1 0.220s 0.009G
denoise_latents 1 2909.754s 0.678G
l2i 1 580.524s 3.008G
img_resize 1 0.317s 0.008G
TOTAL GRAPH EXECUTION TIME: 3929.486s
TOTAL GRAPH WALL TIME: 3929.800s
RAM used by InvokeAI process: 0.05G (-0.297G)
RAM used to load models: 12.92G
VRAM in use: 0.008G
RAM cache statistics:
Model cache hits: 11
Model cache misses: 4
Models cached: 6
Models cleared from cache: 1
Cache high water mark: 9.56/0.00G

 

[osanaiko]

is it really that hard to render on Weak pc ?!

Yes.

GPU : Nvidia GTX 1650

This GPU only has 4gb of VRAM. This is not enough memory to load the model. So the GPU is not used at all...

MY Ram : 16 giga

Your local RAM should be possible load the model + other stuff. Buuuuttt.... doing the crazy number of calculations will take excessive time on a CPU, even the latest ones, compared to a GPU.

I've previously looked into this and as a rough guide, a top-end CPU is roughly 20x slower than a top-end GPU for doing the massively parallel computation needed for Diffusion algorithms (or LLM or Raytrace Rendering).

Given that you have a GTX1650 i'd say that you probably have a 7-10 year old CPU that was not a "top-end" unit in the first place.

So with all that - CPU processing a 64x64 image taking one hour? yes, that sounds believable.

It's sad but true, you need minimum 8GB VRAM GPU (and 12+GB is much better!) to use the local image generation tools.

 

[NoTraceOfLuck]

View attachment 5338599

View attachment 5338600

is it really that hard to render on Weak pc ?!
i tried to render it on 64x64 lowest resolution and i took for me hours to happen

MY Ram : 16 giga
GPU : Nvidia GTX 1650

or do i have anyissue with my setup

Log

If you haven't yet, make sure you enable Low VRAM Mode. There are instructions here:

You must be registered to see the links

Wherever you installed Invoke, you will have a file called ' invokeai.yaml ' and you need to add the one line "enable_partial_loading: true" to it.

I really don't think this will fix the issue, but it's worth a try. 4gb is very low, and the Windows operating system itself uses around 1-2gigs, so you're really running with only around 2gb of VRAM.

You can also look at some SD 1.5 models, these will work on your system. This one for example:

You must be registered to see the links

SD 1.5 is an older model, but takes much less resources.

 

[alesanand]

Is this better than comfyui? I want to make anime characters very close to its original style in different poses like this,
But comfyui seems too complicated. So I would like to know if invoke can give me the result I want but more easily.

 

[Gogameth]

good guide
hacks like me have become superb artists thanks to you

also, is there an alternative to flux for us low vram users? my invoke won't even let me install flux because it runs out of memory

 

[osanaiko]

also, is there an alternative to flux for us low vram users? my invoke won't even let me install flux because it runs out of memory

Blackforestlabs, the creator of flux, have a few public releases on hugging face.

The original Flux1-dev model with no quantization (FP16) is 23.8GB. This would not fit in a 24gb vram card, but would work on a 5090 (!!) or a R6000

They announced that there is also BF16, FP8 and FP4 versions on hugging face, but i could only find a FP8 model in the folders, which is 9.5GB. This would probably work on a 12GB VRAM card and definitely on a 16GB.

There have previously been some even more quantized versions around but I can't find them on Huggingface anymore. I guess BF have chased them down and had them removed.

So basically, there's no hope for Flux on older/low vram cards at the moment.

Actually I found some quantized flux variants:

You must be registered to see the links

So there is a quant all the way down to 4gb: flux1-dev-Q2_K.gguf. Output quality is probably pretty shite though.

Check out the Qwen variants too, they are also very good at natural language -> image generation.

 

[NoTraceOfLuck]

Is this better than comfyui? I want to make anime characters very close to its original style in different poses like this,
But comfyui seems too complicated. So I would like to know if invoke can give me the result I want but more easily.

I wouldn't say it is "better" than ComfyUI. It is a different tool than ComfyUI. ComfyUI is very complicated, but very flexible. Invoke is much simpler, but doesn't support as many options as ComfyUI. If all you want to do is create some cool images, I always recommend Invoke.

Invoke is capable of doing what you want. For example, here is a Nami LoRA:

You must be registered to see the links

It even has different versions for all of her different styles across the different One Piece seasons / movies so you can get the style you want:


There are other LoRAs for pretty much any other character you can think of.


If you follow my tutorial you will learn how to do all of this. Here I just generated this:

 

[NoTraceOfLuck]

good guide
hacks like me have become superb artists thanks to you

also, is there an alternative to flux for us low vram users? my invoke won't even let me install flux because it runs out of memory

Make sure you are installing the one that says "Quantized" and make sure you follow the steps for Low VRAM Mode I mentioned here: https://f95zone.to/threads/your-noob-to-expert-guide-in-making-great-ai-art.256631/post-18362545

They claim Invoke can run Quantized Flux on as low as 6gb of VRAM, but I have never tried this. If you have less than that, I'm not really sure there's a good alternative right now, at least not in Invoke.

 

[Gogameth]

i see, thank you again

i have 8gb of vram, i tested it and it works with 2d images.
as you said i was trying to install a nonquantized flux 20gb model, so perhaps that was the problem

 

[osanaiko]

i see, thank you again

i have 8gb of vram, i tested it and it works with 2d images.
as you said i was trying to install a nonquantized flux 20gb model, so perhaps that was the problem

I made a helpful diagram:

 

[1nsomniac22]

The biggest limitation I'm seeing in the current release of InvokeAI is the lack of support for ControlNet OpenPose editing. You can get around this... but it's a PITA. Even Automatic1111 had a plug in to directly edit the OpenPose model in screen.

Other than that one thing I'm using it exclusively. I've abandoned A1111 and never really got past trying out Comfy UI (yes, as said it's super customizable... and has a super punishing learning curve).