Perfect LoRA Dataset from ONE Photo? The Qwen Revolution in ComfyUI — Transcript

Full transcript

1. 0:00Hey everyone. Today we're kicking off a
2. 0:02series on training a specific module for
3. 0:04our AI. We're talking about training a
4. 0:07Laura file. So why do we even need this
5. 0:09file? It's all about control. We create
6. 0:12a Laura to cut through the chaotic
7. 0:14diffusion randomness and teach the model
8. 0:16to recognize specific visual patterns.
9. 0:18Whether that's a person's face, a unique
10. 0:20architectural style, or complex
11. 0:23technical objects. The goal here is to
12. 0:25lock down the weights in the latent
13. 0:27space, giving us 100% consistency in
14. 0:30every single generation. Without Allora,
15. 0:32you're relying on generative luck. With
16. 0:35Aora, you're the one in charge. If you
17. 0:37want to create an Instagram model who
18. 0:39looks exactly the same, same body, same
19. 0:42face in every photo and video, you need
20. 0:44a data set. A bunch of photos with
21. 0:46specific descriptions for each one. How
22. 0:49do you create this data set? Option one,
23. 0:52find a girl on Instagram and use her
24. 0:54photos as a base. Swap the face on every
25. 0:56single photo and that's it. You're ready
26. 0:58for training. We aren't stealing
27. 1:00anything. We're just training the file
28. 1:02on those shapes and patterns. Just a
29. 1:05tip, try to pick photos without tattoos.
30. 1:07To swap the face, just open my face swap
31. 1:10workflow, drop in the face you want and
32. 1:12the photo from Instagram. Do it 20 times
33. 1:15and you're done. This workflow is
34. 1:17available on my Mimic PC. Just click the
35. 1:19link and it opens right up. Option two,
36. 1:22this is my Q data's workflow. It lets
37. 1:25you take a single photo and turn it into
38. 1:27many different poses and lighting
39. 1:29setups. The main thing is that your
40. 1:31reference photo needs to be prepped with
41. 1:32a face swap first. I won't spend too
42. 1:35much time on the face swap part. Just
43. 1:37plug in the photo and get your result.
44. 1:39Let's talk about Quen instead. Key notes
45. 1:42for the Quen workflow. The base we're
46. 1:45using Quen image edit 2511 in BF-16 but
47. 1:49loading it through FP8. This is crucial
48. 1:51for saving VRAM without losing precision
49. 1:54during token processing. The clip model
50. 1:56here is Quen 2.5 which understands
51. 1:59spatial relationships way better than
52. 2:01the standard SDXL clip. Scaling the
53. 2:04image scale to total pixels node is set
54. 2:06to 1 megapixel. This is the gold
55. 2:09standard for Quen. It automatically fits
56. 2:11the input photo to the right aspect
57. 2:13ratio without stretching the face, which
58. 2:15helps avoid artifacts. Automation. The
59. 2:19heart of the workflow is the CR prompt
60. 2:21list. I've put in 30 presets from 85 mm
61. 2:25Rembrandt lighting portraits to full
62. 2:27body bird's eye shots. Each prompt goes
63. 2:29through a specific Quen image edit plus
64. 2:31encoder, which blends the image and text
65. 2:34at the tensor level for maximum
66. 2:35accuracy. Speed. For speed, we've hooked
67. 2:38up a lightning lora. Check the K
68. 2:41sampler, just six steps, oiler, and the
69. 2:44Beta 57uler.
70. 2:46We're keeping CFG at one. This is
71. 2:48important because lightning models start
72. 2:50to overburn the image at high CFG, and
73. 2:53we need natural skin texture for our
74. 2:54data set. Optimization. If you have less
75. 2:58than 12 GB of VRAM, definitely use Quen
76. 3:01image VAE with a separate loader. It
77. 3:03frees up cache space for the unit. The
78. 3:06noise is set to 1.0 because we're
79. 3:08building new frames from scratch based
80. 3:10on the reference. The big win here is
81. 3:12consistency. You get a set of images
82. 3:14with a clean background and lighting
83. 3:16that ranges from 2,700 Kelvin to 6,500
84. 3:20Kelvin. This is the perfect base for
85. 3:22training a Laura that understands
86. 3:24shadows and anatomy. Final steps for
87. 3:27your data set. All right. Once you have
88. 3:29your 20 to 25 photos of your model in
89. 3:32different poses and moods, you need
90. 3:34descriptions. Without text descriptions
91. 3:36for every photo, training is pointless.
92. 3:39Make sure the file names match the image
93. 3:41names exactly, just with a different
94. 3:43file extension. Take your photos, upload
95. 3:46them to GPT in batches of five, and ask
96. 3:48it to describe them. The most important
97. 3:51part, include a trigger word. This can
98. 3:54be a name or any unique word. Later,
99. 3:56when you're generating images, you'll
100. 3:58put this word in your prompt to activate
101. 4:00the Laura. Pack everything into a folder
102. 4:02like this, and we're ready to train.
103. 4:05I'll give you the full training
104. 4:06instructions in the next video. So,
105. 4:08subscribe to my YouTube and Patreon. I
106. 4:10post all my workflows there with direct
107. 4:12links to the server. You just click and
108. 4:15everything installs in minutes. Links
109. 4:17are in the description.